Data augmentation including background modification using robust prediction with neural networks

Abstract

In various examples, backgrounds of objects can be modified to generate training images. Segmentation masks can be generated and used to generate object images that include image data representing the objects. The object images can be integrated into different backgrounds and used for data augmentation in training of a neural network. Data augmentation can also be performed using tone adjustment (e.g., of the object images) and / or rendering three-dimensional capture data corresponding to the objects from selected views. Inference scores can be analyzed to select backgrounds of images to include in a training dataset. Backgrounds can be selected and training images can be added to the training dataset iteratively during training (e.g., between epochs). Further, early or late fusion using object mask data can be employed to improve inference performed by a neural network trained using object mask data.

Description

BACKGROUND

[0001] When training a neural network to perform a prediction task, such as object classification, the accuracy of the trained neural network is often limited by the quality of the training dataset. In order to train a robust neural network, the network should be trained using challenging training images. For example, in training a neural network for hand pose recognition (e.g., thumbs up, peace sign, fist, etc.), the network can have difficulty detecting hand poses in front of certain environmental features. If the pose includes extended fingers, the network can perform well when the pose is in front of a mostly solid color environment, but can perform poorly when the pose is in front of an environment that includes certain color patterns. As another example, the network can have difficulty with certain poses from certain angles or when the environment and hand have similar color tones.

[0002] However, whether a particular training image is challenging for a neural network can depend on many factors, such as the prediction task being performed, the architecture of the neural network, and other training images seen by the network. Thus, it is difficult to construct a training dataset by anticipating which training images should be used to train the network that will produce a robust trained network. It is possible to estimate which features of training images can be challenging for a neural network. However, even if such an estimate is possible and accurate, it can not be possible or practical to obtain enough images exhibiting those features to sufficiently train the network. SUMMARY

[0003] Embodiments of the present disclosure relate to data augmentation including background filtering for robust predictions using neural networks. Systems and methods are disclosed that provide data augmentation techniques, such as those based on background filtering, that can be used to increase the robustness of a trained neural network.

[0004] In contrast to conventional systems, the present disclosure provides for modifying the background of an object to generate training images. Segmentation masks can be generated and used to generate object images that include image data representing an object. The object images can be integrated into different backgrounds and used for data augmentation in training a neural network. Other aspects of the present disclosure provide for data augmentation using tone adjustment (e.g., of object images) and / or rendered three-dimensional capture data corresponding to an object from a selected view direction. The present disclosure also provides for analyzing inference scores to select backgrounds of images to include in a training dataset. The backgrounds can be selected during training (e.g., between epochs) and training images can be iteratively added to the training dataset. Further, the present disclosure provides for early or late fusion using object mask data to improve inference performed by a neural network trained using the object mask data. BRIEF DESCRIPTION OF DRAWINGS

[0005] Systems and methods for data augmentation including background filtering using robust prediction with neural networks are described in detail below with reference to the accompanying drawings, wherein:

[0006] FIG. 1 is a data flow diagram illustrating an example process for training one or more machine learning models based at least on integrating object images with backgrounds in accordance with some embodiments of the present disclosure;

[0007] FIG. 2 is a data flow diagram illustrating an example process for generating object images and integrating object images with one or more backgrounds in accordance with some embodiments of the present disclosure;

[0008] FIG. 3 includes an example of preprocessing that can be used to generate object masks for generating object images in accordance with some embodiments of the present disclosure;

[0009] FIG. 4 is an illustration of how a three-dimensional capture of an object can be rasterized from multiple views in accordance with some embodiments of the present disclosure;

[0010] FIG. 5A is a data flow diagram illustrating an example of inference using early fusion of machine learning models and object mask data in accordance with some embodiments of the present disclosure;

[0011] FIG. 5B is a data flow diagram illustrating an example of inference using late fusion of machine learning models and object mask data in accordance with some embodiments of the present disclosure;

[0012] FIG. 6 is a flow diagram illustrating a method for training one or more machine learning models based at least on integrating object images with at least one background in accordance with some embodiments of the present disclosure;

[0013] FIG. 7 is a flow diagram illustrating a method of inference using machine learning models in accordance with some embodiments of the present disclosure, where the input corresponds to a mask of an image and at least a portion of the image;

[0014] FIG. 8 is a flow diagram illustrating a method for selecting backgrounds of objects for training one or more machine learning models in accordance with some embodiments of the present disclosure;

[0015] FIG. 9 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure;

[0016] FIG. 10 is a block diagram of an example data center suitable for implementing some embodiments of the present disclosure;

[0017] FIG. 11A These are illustrations of example autonomous vehicles according to some embodiments of the present disclosure;

[0018] FIG. 11B According to some embodiments of this disclosure FIG. 11A Examples of camera positions and fields of view for autonomous vehicles;

[0019] FIG. 11C According to some embodiments of this disclosure FIG. 11A A block diagram of an example system architecture for an example autonomous vehicle; and

[0020] FIG. 11D It is a cloud-based server and according to some embodiments of this disclosure FIG. 11A Here is a system diagram illustrating communication between example autonomous vehicles. Detailed Implementation

[0021] Systems and methods related to data augmentation, including background filtering for robust prediction using neural networks, are disclosed. Embodiments of this disclosure relate to data augmentation including background filtering for robust prediction using neural networks. Systems and methods for providing data augmentation techniques, such as those based on background filtering, are disclosed that can be used to increase the robustness of trained neural networks.

[0022] The disclosed embodiments can be implemented using a variety of different systems, such as automotive systems, robotics, aviation systems, medical systems, boating systems, smart area monitoring systems, simulation systems, and / or other technical fields. The disclosed methods can be used for any perception-based or more generally image-based analysis using machine learning models, such as for the monitoring and / or tracking of objects and / or environments.

[0023] Applications of the disclosed technology include multi-mode sensor interfaces, which can be used in healthcare settings. For example, intubated or otherwise unable to communicate verbally can use gestures or postures interpreted by a computing system. Applications of the disclosed technology also include autonomous driving and / or vehicle control or interaction. For example, the disclosed technology can be used to implement [equipment / systems / etc.]. FIG. 11A-11D The vehicle 1100 uses in-cabin sensor to recognize hand gestures or hand positions to control convenience functions, such as multimedia options. Gesture recognition can also be applied to the external environment of the vehicle 1100 to control any of a variety of autonomous driving control operations, including advanced driver assistance systems (ADAS) functions.

[0024] As various examples, the disclosed technology can be implemented in a system that includes or is included in one or more of a system for performing conversational AI or personal assistant operations, a system for performing simulation operations, a system for performing simulation operations to test or validate autonomous machine applications, a system for performing deep learning operations, a system implemented using edge devices, a system incorporating one or more virtual machines (VMs), a system implemented at least in part in a data center, or a system implemented at least in part using cloud computing resources.

[0025] In contrast to conventional systems, the present disclosure provides for identifying a region in an image that corresponds to an object and using the region to filter, remove, replace, or otherwise modify a background of the object and / or the object itself to generate a training image. In accordance with the present disclosure, a segmentation mask can be generated that identifies one or more segments that correspond to the object and one or more segments that correspond to a background of the object in a source image. The segmentation mask can be applied to the source image to identify the region that corresponds to the object, for example, to generate an object image that includes image data representative of the object. The object image can be integrated into different backgrounds and used for data augmentation in training a neural network. Other aspects of the present disclosure provide for data augmentation using tone adjustment (e.g., of the object image) and / or rendering of three-dimensional captured data that corresponds to the object from a selected view direction.

[0026] Further aspects of the present disclosure provide for methods for selecting backgrounds of objects for training a neural network. In accordance with the present disclosure, a machine learning model (MLM) can be trained at least in part and inference data can be generated by the MLM using images that include different backgrounds. The MLM can include a neural network during training or a different MLM. Inference scores corresponding to the inference data can be analyzed to select one or more features of the training images, such as particular backgrounds or types of backgrounds of images to include in a training dataset. The images can be selected from existing images or generated using any suitable method, such as those described herein, using object images and backgrounds. In at least one embodiment, one or more features can be selected and one or more corresponding training images can be iteratively added to the training dataset during training.

[0027] The present disclosure also provides methods of using object mask data to improve inference performed by a neural network trained using object mask data. Post-fusion can be performed where one set of inference data is generated from a source image and another set of inference data is generated from an image that captures object mask data, e.g., an object image (e.g., using two copies of a neural network). The sets of inference data can be fused and used to update the neural network. In further examples, early fusion can be performed where a source image and an image that captures object mask data are combined and inference data is generated from the combined image. The object mask data can be used to attenuate or otherwise modify the background of the source image.

[0028] Reference is now made to FIG. 1 , FIG. 1 is a dataflow diagram illustrating an example process 100 for training one or more machine learning models based at least on integrating object images with backgrounds, in accordance with some embodiments of the present disclosure. The process 100 is described with respect to a machine learning model (MLM) training system 140, for example. Among other potential components, the MLM training system 140 can include a background integrator 102, an MLM trainer 104, an MLM post-processor 106, and a background selector 108.

[0029] At a high level, the process 100 can include the background integrator 102 receiving one or more backgrounds 110 (which can be referred to as background images) and object images 112 corresponding to one or more objects (e.g., to be classified, analyzed, and / or detected by the MLMs 122). The background integrator 102 can integrate the object images 112 with the backgrounds 110 to produce image data that captures (e.g., represents) the backgrounds 110 and at least a portion of the objects in one or more images. The MLM trainer 104 can produce inputs 120 for one or more MLMs 122 from the image data. The MLMs 122 can process the inputs 120 to generate one or more outputs 124. The MLM post-processor 106 can process the outputs 124 to generate prediction data 126 (e.g., inference scores, object class labels, object bounding boxes or shapes, etc.). The background selector 108 can analyze the prediction data 126 and select at least one or more of the backgrounds 110 and / or the objects for training based on the prediction data 126. In some embodiments, the process 100 can repeat for any number of iterations until the one or more MLMs 122 are sufficiently trained or the background selector 108 can be used for selecting the backgrounds 110 for a first training iteration and / or any other iteration, either once or intermittently.

[0030] For example, and without limitation, the MLMs 122 described herein can include any type of machine learning model, such as machine learning models that use linear regression, logistic regression, decision trees, support vector machines (SVMs), Naive Bayes, k-nearest neighbors (Knn), K-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, neural networks (e.g., autoencoders, convolutional, recurrent, perceptrons, long / short-term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolutional, generative adversarial, liquid state machines, etc.), and / or other types of machine learning models.

[0031] The process 100 can be used, at least in part, to train one or more MLMs 122 to perform a prediction task. The present disclosure focuses on gesture recognition and / or hand gesture recognition, and more specifically hand pose recognition. However, the disclosed techniques are widely applicable to training MLMs to perform a variety of possible prediction tasks, such as image and / or object classification tasks. Examples include object detection, bounding box or shape determination, object classification, pose classification, hand gesture classification, and so on. For example, FIG. 2 An example of an image 246 that can be captured by the input 120 to the MLMs 122 is shown. The process 100 can be used to train the MLMs 122 to predict the pose of a hand depicted in the image 246 (e.g., thumb up, thumb down, fist, peace sign, open hand, OK sign, etc.).

[0032] In some embodiments, one or more iterations of the process 100 can not include training one or more MLMs 122. For example, the background selector 108 can use an iteration to select one or more backgrounds 110 and / or objects corresponding to the object images 112 for inclusion in a training dataset used by the MLM trainer 104 for training. Further, in some examples, the process 100 can use one MLM 122 to select backgrounds 110 and / or objects for a training dataset in an iteration, and can use the training dataset to train the same or a different MLM 122 (e.g., in a subsequent iteration of the process 100 or otherwise). For example, the MLM 122 used to select from the backgrounds 110 for training can be partially or fully trained to perform a prediction task. In cases where an iteration of the process 100 uses a trained or partially trained MLM 122 to select from the backgrounds 110 for training another MLM 122, the iteration can be used to guide the training of the other MLM 122 by selecting challenging background images and / or background and object image combinations for training.

[0033] In various examples, one or more iterations of the process 100 can form a feedback loop in which the background selector 108 uses prediction data of an iteration (e.g., a training epoch) to select one or more backgrounds 110 and / or objects corresponding to the object images 112 to include in a subsequent training dataset. In a subsequent iteration (e.g., a subsequent training epoch) of the process 100, the background integrator 102 can generate or otherwise prepare or select corresponding images that the MLM trainer 104 can incorporate into a training dataset. The training dataset can then be applied to the MLM 122 being trained to generate prediction data 126 used by the background selector 108 to select one or more backgrounds 110 and / or objects corresponding to the object images 112 to include in a subsequent training dataset. The feedback loop can be used to determine enhancements to ensure continued improvement in accuracy, generalization performance, and robustness of the MLM 122 being trained.

[0034] In various examples, based at least on the background selector 108 selecting one or more backgrounds and / or objects, the MLM trainer 104 incorporates one or more images from the background integrator 102 that include the selected backgrounds and / or the selected backgrounds and object combinations. As an example, the MLM trainer 104 can add the one or more images to a training dataset used for a previous training iteration and / or epoch. The training dataset can grow for each iteration. However, in some cases, the MLM trainer 104 can also remove one or more images from the training dataset used for a previous training iteration and / or epoch (e.g., based on the selection by the background selector 108 for removal and / or based on the training dataset exceeding a threshold number of images). In the illustrated example, the background integrator 102 can generate the one or more images to include in the training dataset at the beginning of an iteration of the process 100 based on the selection made by the background selector 108. In other examples, the one or more images can be pre-generated by the background integrator 102, such as at least partially prior to any training using the process 100 and / or during one or more previous iterations. In the case of pre-generated images, the MLM trainer 104 can retrieve the pre-generated images from storage based on the selection made by the background selector 108.

[0035] The selection of the background described herein can refer to selecting a background to include for training. The selection of the background can also include selecting an object to include in the image with the background. In at least one embodiment, the background selector 108 can select a background and / or a background and object combination based at least on a confidence of the MLM 122 in one or more predictions made using the MLM 122. In various examples, the confidence can be captured by a set of inference scores that correspond to predictions of one or more prediction tasks performed by the one or more MLMs 122 on one or more images. For example, the predictions can be made in the current iteration and / or one or more previous iterations of the process 100. An inference score can refer to a score that the MLM is trained to provide or is being trained to provide with respect to a prediction task or a portion thereof. In some examples, the inference score can represent a confidence of the MLM with respect to one or more respective outputs 124 (e.g., tensor data) and / or can be used to determine or compute a confidence with respect to the prediction task. For example, the inference score can represent a confidence of the MLM 122 in an object detected in an image belonging to a target class (e.g., a probability that the input 120 belongs to the target class).

[0036] The background selector 108 can use a variety of possible methods to select one or more of the backgrounds and / or background and object combinations based on the inference scores. In at least one embodiment, the background selector 108 can select one or more particular backgrounds and / or objects based at least on an analysis of inference scores corresponding to images containing those elements. In at least one embodiment, the background selector 108 can select one or more backgrounds and / or objects of a particular class or type or having one or more other particular characteristics based at least on an analysis of inference scores of images corresponding to those elements including those one or more characteristics (e.g., particular background, particular object, texture, color, lighting conditions, object and / or overlay, hue, perspective, direction, skin tone, size, theme, included background elements, etc.). FIG. 4 For example, the background selector 108 can select at least one background including a louver and a gesture including an open palm based at least on inference scores of images sharing those characteristics. As another example, the background selector 108 can select a particular background based at least on inference scores of images including that background. As an additional example, the background selector 108 can select a particular background and an object class of object (e.g., thumbs up, thumbs down, etc.) based at least on inference scores of images including that background and the object class of object.

[0037] In some cases, the inference score can be evaluated by the background selector 108 based at least on computing a confusion score. The confusion score can be used as a measure of quantifying the relative network confusion with respect to predictions made on one or more images having a set of particular characteristics. In cases where the confusion score exceeds a threshold value (e.g., indicating sufficient confusion), the background selector 108 can select at least some elements of the set of characteristics (e.g., a background or a combination of a background and an object) to modify the training dataset. In at least one embodiment, the confusion of a background can be based at least in part on a number of correct and incorrect predictions made on images including elements having a set of characteristics. For example, the confusion score can be computed based at least on a ratio between correct and incorrect predictions. Additionally or alternatively, the confusion score can be computed based at least on a difference in prediction accuracy or inference scores of images having elements of a set of characteristics (e.g., indicating when to use a particular background or background type, there is a large difference between target classes).

[0038] The background selector 108 can select one or more backgrounds and / or background and object combinations to include in at least one image of the training dataset based at least on selecting one or more different element characteristics. For example, the background selector 108 can select one or more different element characteristics based at least on a corresponding confusion score. The background selector 108 can rank different sets of characteristics and select one or more sets based on the ranking for modifying the training dataset. As a non-limiting example, the background selector 108 can select the top N particular backgrounds or background and object combinations (or other sets of characteristics), where N is an integer (e.g., for each confusion score that exceeds a threshold value).

[0039] The background integrator 102 can select, retrieve (e.g., from storage), and / or generate one or more images that satisfy the selection made by the background selector 108 to modify the training dataset. In cases where the background integrator 102 generates an image from a selected background, the background integrator 102 can include an entire background image or one or more portions of the background in the image. For example, the background integrator 102 can sample a region from the background (e.g., a rectangle of a size determined based on the input 120 of the MLM 122) using a random or non-random sampling method. Thus, the image processed by the MLM 122 can include an entire background image or a region of the background image (e.g., a sampled region). Similarly, in cases where the background integrator 102 generates an image from a selected object, the background integrator 102 can include an entirety of an object image or a portion of the object image in the image.

[0040] In at least one embodiment, the background integrator 102 can generate one or more synthetic backgrounds. For example, synthetic backgrounds can be generated (either empirically measuring network performance or intuitively) for situations to understand network bias and sensitivity. For example, synthetic backgrounds of particular types (e.g., dots and stripes) can be generated for networks that are sensitive to those particular textures, where the background selector selects that background type. One or more synthetic backgrounds can be generated prior to any training of the MLM 122 and / or during training (e.g., between iterations or epochs). Synthetic backgrounds can be generated using a variety of possible methods, for example, based at least on rendering a three-dimensional virtual environment associated with a background type, generating a texture including a selected pattern by algorithm, modifying an existing background or image, etc.

[0041] According to aspects of the present disclosure, an object image 112 can be extracted from one or more source images, and the background integrator 102 can replace or modify the original background of the source image using one or more backgrounds 110. Referring now to FIG. 2 , FIG. 2 is a dataflow diagram illustrating an example process 200 for generating an object image 212 and integrating the object image 212 with one or more backgrounds 110, according to some embodiments of the present disclosure.

[0042] As an example, the process 200 is described with respect to an object image extraction system 202. Among other potential components, the object image extraction system 202 can include a region identifier 204, a preprocessor 206, and an image data determiner 208.

[0043] As an overview, in the process 200, the region identifier 204 can be configured to identify a region within a source image. For example, the region identifier 204 can identify a region 212A within a source image 220 that corresponds to an object (e.g., a hand) having a background in the source image 220. The region identifier 204 can further generate a segmentation mask 222 including a segment 212B that is based on identifying that the region 212A corresponds to the object. The region identifier 204 can further detect a location of the object to define a region 230 of the source image 220 and / or the segmentation mask 222. The preprocessor 206 can process at least a portion of the segment 212B in the region 230 of the segmentation mask 222 to produce an object mask 232. The image data determiner 208 can generate an object image 212 from the source image 220 using the object mask 232. The object image can then be provided to the background integrator 102 for integration with one or more backgrounds 110 (e.g., overlaying or superimposing the object on a background image).

[0044] According to various embodiments, one or more object images 112, such as object images 212, may be generated before and / or during training MLM 122. For example, one or more object images 112 may be generated (e.g., as...) FIG. 2 As described, the images are stored and then retrieved as needed to generate input 120 in process 100. As another example, one or more object images 112 may be generated during process 100, for example, immediately or as needed by the background integrator 102. In some embodiments, object images 112 are generated immediately in process 100 and may then be stored and / or reused and / or used later to train MLMs other than MLM 122 in subsequent iterations of process 100.

[0045] As described herein, region identifier 204 can identify region 212A within source image 220, which corresponds to an object (e.g., a hand) having a background in source image 220. In the example shown, region identifier 204 can also identify region 210A within source image 220 that corresponds to the background of an object. In other examples, region identifier 204 can identify only region 212A.

[0046] In at least one embodiment, region identifier 204 can identify region 212A to determine at least segment 212B of source image 220 corresponding to an object, at least based on image segmentation performed on the source image. Image segmentation can also be used to identify region 210A to determine segment 212B of source image 220 corresponding to a background, at least based on image segmentation performed on the source image 220. In at least one embodiment, region identifier 204 can generate data representing a segmentation mask 222 from source image 220, wherein segmentation mask 222 indicates the segment corresponding to an object ( FIG. 2 The white pixels in the middle) segment 212B and / or the segment 210B corresponding to the background ( FIG. 2 (The black pixels in the image).

[0047] Region identifier 204 can be implemented in a variety of possible ways, such as using AI-driven background removal. In at least one embodiment, region identifier 204 includes one or more MLMs trained to classify or label individual pixels or groups of pixels in an image. For example, the MLM can be trained to identify foreground (e.g., corresponding to an object) and / or background in an image, and image fragments can correspond to foreground and / or background. As a non-limiting example, region identifier 204 can be implemented using NVIDIA's RTX Greenscreen background removal technology. In some examples, the MLM can be trained to identify object types and label pixels accordingly.

[0048] In at least one embodiment, region identifier 204 can include one or more object detectors, such as an object detector trained to detect an object (e.g., a hand) that will be classified by MLM 122. An object detector can be implemented using one or more MLMs trained to detect an object. An object detector can output data indicative of an object’s location and can be used to define a region 230 of source image 220 and / or segmentation mask 222 that includes the object. For example, an object detector can be trained to provide a bounding box or shape of an object, and the bounding box or shape can be used to define region 230.

[0049] In the illustrated example, region 230 can be defined by expanding a bounding box, while in other examples, a bounding box can be used as region 230. In this example, region identifier 204 can identify a segment 212B that corresponds to an object of source image 220 by applying to the MLM. In other examples, region identifier 204 can apply region 230 to the MLM instead of (or in addition to, in some embodiments) the source image. By applying source image 220 to the MLM, the MLM can have additional context that is not available in region 230, which can improve the accuracy of the MLM.

[0050] In embodiments where region 230 is determined, preprocessor 206 can perform preprocessing based at least on region 230. For example, region 230 of segmentation mask 222 can be preprocessed by preprocessor 206 before being used by image data determiner 208. In at least one embodiment, preprocessor 206 can crop image data from segmentation mask 222 that corresponds to region 230 and process the cropped image data to produce object mask 232. Preprocessor 206 can perform various types of preprocessing on region 230, which can improve the capabilities of image data determiner 208.

[0051] Reference is now made to FIG. 3 , FIG. 3 includes an example of preprocessing that can be used to generate an object mask for generating an object image in accordance with some embodiments of the present disclosure. By way of example, preprocessor 206 can crop segmentation mask 222, resulting in object mask 300A. Preprocessor 206 can perform dilation on object mask 300A, resulting in object mask 300B. Preprocessor 206 can then blur object mask 300B, resulting in object mask 232. Object mask 232 can then be used by image data determiner 208 to generate object image 212.

[0052] The preprocessor 206 can dilate the segment 212B corresponding to the object mask 300A using dilation. For example, the segment 212B can be dilated to the segment 210B corresponding to the background. In embodiments, the preprocessor 206 can perform binary dilation. Other types of dilation can be performed, such as grayscale dilation. As an example, the preprocessor 206 can first blur the object mask 300A and then perform grayscale dilation. Dilation can be used to increase the robustness of the object mask 232 to errors in the segmentation mask 222. For example, in cases where the object includes a hand, the palm of the hand can sometimes be classified as being in the background. Dilation is one approach to fix such potential errors. Other mask preprocessing techniques are within the scope of the present disclosure, such as binary or grayscale erosion. As an example, erosion can be performed on the segment 210B corresponding to the background.

[0053] The preprocessor 206 can use blurring (e.g., Gaussian blurring) to help the background integrator 102 soften the transition between image data corresponding to the object in the object image 212 and image data corresponding to the background 110. To soften the transition between the object and the background, the area corresponding to the edges of the object mask 232 can be sharp and artificial. Using a blending technique such as blurring and then applying the object mask can result in a more natural or realistic transition between the object and the background 110 in the image 246. While mask processing has been described as being performed on the object mask before applying the mask, in other examples the image data determiner 208 can perform similar or different image processing operations when applying the object mask (e.g., to the source image 220).

[0054] Returning to FIG. 2 The image data determiner 208 can generally use an object mask, such as the object mask 232, to generate the object image 212. For example, the image data determiner 208 can use the object mask 232 to identify and / or extract the region 242 from the source image 220 corresponding to the object. In other embodiments, an object mask can not be employed and another technique can be used to identify and / or extract the region 242. When using the object mask 232, the image data determiner 208 can multiply the object mask 232 with the source image 220 to obtain the object image 212, which includes image data representing the region 242 corresponding to the object (e.g., the foreground of the source image 220).

[0055] In integrating the object image 212 with the background 110, the background integrator 102 can use the object image 212 as a mask, and can apply an inverse of the mask to the background 110, resulting in an image that blends the target image. For example, the background integrator 102 can perform alpha blending between the object image 212 and the background 110. Blending the object image 212 with the background integrator 102 can use a variety of potential blending techniques. In some embodiments, the background integrator 102 can use alpha blending to integrate the object image 212 with the background 110. When combining the object image 212 with the background 110, alpha blending can zero out the background from the object image 212, and foreground pixels can be overlaid on the background 110 to generate the image 246, or the pixels can be weighted (e.g., from 0 to 1) when combining image data from the object image 212 and the background 110 according to a blur or other manner applied by the preprocessor 206. In at least one embodiment, the background integrator 102 can employ one or more seamless blending techniques. Seamless blending techniques can aim to produce a seamless boundary between the object in the image 246 and the background 110. Examples of seamless blending techniques include gradient domain blending, Laplacian pyramid blending, or Poisson blending.

[0056] Further examples of data augmentation techniques

[0057] As described herein, the process 200 can be used to augment a training dataset used to train an MLM, such as the MLM 122 using the process 100. The present disclosure provides further methods that can be used to augment a training dataset. According to at least some embodiments, the hue of an object identified in a source image (e.g., the source image 220) can be modified for data augmentation. As an example, where the object represents at least a portion of a human, the skin tone, hair color, and / or other hue can be modified to augment the training dataset. For example, the hue of one or more portions of the region corresponding to the object can be transformed (e.g., uniformly or otherwise). In at least one embodiment, the hue can be selected randomly or non-randomly. In some cases, the hue can be selected based on analysis of the prediction data 126. For example, the hue can be used as a feature for selecting or generating one or more training images (e.g., by the background integrator 102), as described herein.

[0058] Certain regions, such as a background or non-primary or secondary portion of a region, can tend to have a consistent color tone for different real-life changes to the object, which can preserve the original color tone. For example, if the object is a car, the color tone of the panels can change while maintaining the color tone of the lights, bumpers, and tires. In at least one embodiment, the background integrator 102 can perform color tone modification. For example, color tone modification can be performed on one or more portions of the object represented in the object image 212. In other examples, the object image 212 or the object mask 232 can be used (e.g., by the image data determiner 208) to identify image data representing the object and modify the color tone in one or more regions of the source image 220. These examples can not include the background integrator 102.

[0059] According to at least some embodiments, the source image 220 can be rendered from various different views of the object in an environment for data augmentation. Referring now to FIG. 4 , FIG. 4 is an illustration of how a three-dimensional (3D) capture 402 of an object according to some embodiments of the present disclosure can be rasterized from multiple views. In at least one embodiment, the object image extraction system 202 can select a view of the object in the environment. For example, the object image extraction system 202 can select from the views 406A, 406B, 406C, or any arbitrary view of the object in the environment 400. The object image extraction system 202 can then generate the source image 220 based at least on the rasterization of the 3D capture of the object from the view of the environment. For example, the three-dimensional (3D) capture 402 can include depth information captured by a physical or virtual depth sensing camera in a physical or virtual environment (which can be different from the environment 400). In one or more embodiments, the 3D capture 402 can include a point cloud capturing at least a portion of the object and potentially additional elements of the environment 400. For example, if the view 406A is selected, the object image extraction system 202 can use at least the 3D capture 402 to rasterize the source image 220 from the view 406A of the camera 404. In at least one embodiment, the view can be selected randomly or non-randomly. In some cases, the view can be selected based on analysis of the prediction data 126. For example, the view can be used as a feature for selecting or generating one or more training images (e.g., by the background integrator 102), as described herein (e.g., in connection with object classes). In at least one embodiment, the object can be rasterized from the view to generate the object image 112, which can then be integrated with one or more backgrounds using the methods described herein. In other examples, the object can be rasterized with a background 110 (a two-dimensional image) or with other 3D content of the environment 400 to form a background.

[0060] Example of inference using object masks

[0061] As described herein, object masks can be used for data augmentation in training the MLM 122, e.g., using the process 100. In at least one embodiment, the MLM 122 trained using masked data can perform inference on images without utilizing object masks. For example, the input 120 to the MLM 122 during deployment can correspond to one or more images captured by a camera. In such examples, the object masks can only be used for data augmentation. In other embodiments, the object masks can also be utilized for inference. An example of how object masks can be utilized for inference will be described with reference to FIG. 5A and 5B

[0062] Referring now to FIG. 5A , FIG. 5A is a dataflow diagram 500 illustrating an example of inference using early fusion of the MLM 122 and object mask data, in accordance with some embodiments of the present disclosure. In the example of the dataflow diagram 500, the MLM 122 can be trained to perform inference on an image 506 while utilizing an object image 508 of an object mask corresponding to the image 506. For example, the input 120 can be generated from a combination of the image 506 and the object image 508, and then provided to the MLM 122 (e.g., a neural network), which can generate the output 124 including inference data 510. In the case where the MLM 122 includes a neural network, the inference data 510 can include tensor data from the neural network. Post-processing can be performed on the inference data 510 to generate the prediction data 126. FIG. 5A

[0063] The object image 508 is one example of object mask data that can be combined with the image 506 for inference. In the case where early fusion of object mask data is used for inference, as in the dataflow diagram 500, the input 120 to the MLM 122 can similarly be generated during training (e.g., in the process 100). In general, the object mask data can capture information about the way the region identifiers 204 generate masks from source images. By utilizing the object mask data during training and inference, the MLM 122 can learn to interpret any errors or unnatural artifacts that can result from the generation of the object masks. The object mask data can also capture information about the way the preprocessor 206 pre-processes the object masks to capture any errors or unnatural artifacts that can result from the pre-processing or remain after the pre-processing.

[0064] The object image 508 can be generated (e.g., at inference time) using an object image extraction system 202 similar to the object image 212. While the object image 508 is illustrated in the dataflow diagram 500 as being generated from the image 506, in other embodiments, the object image 508 can be generated from a different image, e.g., a different image captured by the camera. FIG. 5A and FIG. 5B ​​In other examples, in addition to or instead of object image 508 (before or after preprocessing by preprocessor 206), segmentation mask 222 and / or object mask 232 can be used.

[0065] Input 120 can be generated from the combination of image 506 and object image 508 (more generally, object mask data) using various methods. In at least one embodiment, image 506 and object image 508 are provided as separate inputs 120 to MLM 122. As a further example, image 506 and object image 508 can be combined to form a combined image and input 120 can be generated from the combined image. In at least one embodiment, object mask data can be used to fade, dim, mark, indicate, distinguish, or otherwise modify one or more portions of image 506 representing background relative to the object or foreground of image 506 as captured by the object mask data. For example, image 506 can be blended with object image 508, resulting in the background of image 506 being faded, dimmed, or defocused (e.g., using a depth of field effect). When combining image 506 and object image 508, the weight used to determine the resulting pixel color can decrease (e.g., exponentially) with distance from the object, as indicated by the object mask data (e.g., using a focus effect).

[0066] Reference is now made to FIG. 5B , FIG. 5B is a dataflow graph 502 illustrating an example of inference using late fusion of MLM 122 and object mask data, in accordance with some embodiments of the present disclosure.

[0067] In FIG. 5BIn the illustrated example, the MLM 122 can provide separate outputs 124 for the image 506 and the object image 508. The outputs 124 can include inference data 510A corresponding to the image 506 and inference data 510B corresponding to the object image 508. Further, the MLM 122 can include separate inputs 120 for the image 506 and the object image 508. For example, the MLM 122 can include multiple copies of the MLM 122 trained to perform inference in images, where one copy performs inference on the image 506 and generates inference data 510A and another copy performs inference on the object image 508 and generates inference data 510B (e.g., in parallel). Post-processing can be performed on the inference data 510A and the inference data 510B and late fusion can be used to generate the prediction data 126. For example, corresponding tensor values across the inference data 510A and the inference data 510B can be combined (e.g., averaged) to fuse the inference data, and further post-processing can be performed on the fused inference data to generate the prediction data 126. In at least one embodiment, tensor values of the inference data 510A and the inference data 510B can be combined using weights (e.g., using a weighted average). In at least one embodiment, the weights can be adjusted on a validation dataset.

[0068] Inference using the MLM 122 can also include temporal filtering of inference scores to generate the prediction data 126, which can improve temporal stability of predictions. Further, the illustrated examples are primarily related to static pose recognition. However, the disclosed techniques can also be applied to dynamic pose recognition, which can be referred to as gesture. To train and use an MLM to predict a pose, in at least one embodiment, multiple images can be provided to the MLM 122 that capture an object over a period of time or number or sequence of frames. In cases where object mask data is used, object mask data can be provided for each input image.

[0069] Referring now to FIG. 6 Each block of the method 600, and other methods described herein, comprises a computing process that can be performed using any combination of hardware, firmware, and / or software. For instance, various functions can be carried out by a processor executing instructions stored in memory. The methods can also be embodied as computer-usable instructions stored on computer storage media (e.g., both memory and storage media). The methods can be provided via a standalone application, a service or hosting service (standalone or in combination with another hosted service), or a plug-in to another product, to name a few. FIG. 1 The method 600 is described with respect to the system 140 of FIG. 2 and the system 202 of However, the method can additionally or alternatively be performed by any one system or any combination of systems, including but not limited to those described herein.

[0070] FIG. 6This is a flowchart illustrating a method 600 for training one or more machine learning models based at least on combining an object image with at least one background, according to some embodiments of the present disclosure. In block B602, method 600 includes identifying regions in a first image corresponding to objects having a first background. For example, region identifier 204 may identify region 212A in source image 220 corresponding to an object having a background in source image 220.

[0071] In box B604, method 600 includes determining image data representing an object based at least on a region. For example, image data determiner 208 may determine image data representing an object based at least on region 212A of the object. In at least one embodiment, image data determiner 208 may use object mask 232 or a non-mask-based method to determine the image data.

[0072] In box B606, method 600 includes generating a second image including an object with a second background using image data. Background integrator 102 may generate an image 246 including the object with background 110 based at least on combining the object with background 110 using image data. For example, image data determiner 208 may incorporate image data into object image 212 and provide object image 212 to background integrator 102 for integration with background 110.

[0073] In box B608, method 600 includes training at least one neural network to perform a prediction task using a second image. For example, MLM trainer 104 can use image 246 to train an MLM to classify objects in the image.

[0074] Now for reference FIG. 7 , FIG. 7 This is a flowchart illustrating a method 700 for inference using a machine learning model according to some embodiments of the present disclosure, wherein the input corresponds to a mask of an image and at least a portion of the image. In block B702, method 700 includes obtaining (or accessing) at least one neural network trained to perform a prediction task on the image using input generated from the mask corresponding to an object. For example, it may obtain (access) FIG. 5A Or 5B MLM 122 and may have been based on FIG. 1 The process involved training 100.

[0075] In box B704, method 700 includes generating a mask corresponding to an object in an image, wherein the object has a background in the image. For example, region identifier 204 may generate a segmentation mask 222 corresponding to an object in source image 220, wherein the object has a background in source image 220.

[0076] At block B706, the method 700 includes generating an input to the at least one neural network using the mask. For example, FIG. 5A Or the 5B input 120 can be generated using the segmentation mask 222 (or without using object mask data). The input 120 can capture the object with at least a portion of the background.

[0077] At block B708, the method 700 includes generating at least one prediction of a prediction task based at least on applying the input to the at least one neural network. For example, the MLM 122 can be used to generate at least one prediction of a prediction task based at least on applying the input 120 to the MLM 122, and the output 124 from the MLM 122 can be used to determine the prediction data 126.

[0078] Reference is now made to FIG. 8 , FIG. 8 is a flowchart illustrating a method 800 for selecting a background for training one or more machine learning models of objects, in accordance with some embodiments of the present disclosure. At block B802, the method 800 includes receiving an image of one or more objects having a plurality of backgrounds. For example, the MLM training system 140 can receive an image of one or more objects having a plurality of backgrounds 110.

[0079] At block B804, the method 800 includes generating a set of inference scores corresponding to a prediction task using the image. For example, the MLM trainer 104 can provide the input 120 to one or more MLMs 122 (or different MLMs) to generate the output 124, and the MLM post-processor 106 can process the output 124 to produce the prediction data 126.

[0080] At block B806, the method 800 includes selecting a background based at least on the one or more inference scores. For example, the background selector 108 can select one or more of the backgrounds 110 based at least on the prediction data 126.

[0081] At block B808, the method 800 includes generating an image based at least on integrating the object with the background. For example, the background integrator 102 can generate an image based at least on integrating the object with the background (e.g., using the object image 112 and the background 110).

[0082] At block B810, the method 800 includes training at least one neural network to perform a prediction task using the image. For example, the MLM trainer 104 can train one or more MLMs 122 using the image.

[0083] Example computing device

[0084] FIG. 9is a block diagram of an example computing device 900 suitable for implementing some embodiments of the present disclosure. The computing device 900 can include an interconnection system 902 coupling the following components: a memory 904, one or more central processing units (CPU) 906, one or more graphics processing units (GPU) 908, a communication interface 910, an input / output (I / O) port 912, an input / output component 914, a power supply 916, one or more presentation components 918 (e.g., a display), and one or more logic units 920. In at least one embodiment, the computing device 900 can include one or more virtual machines (VMs), and / or any component thereof can include a virtual component (e.g., a virtual hardware component). For a non-limiting example, the one or more GPUs 908 can include one or more vGPUs, the one or more CPUs 906 can include one or more vCPUs, and / or the one or more logic units 920 can include one or more virtual logic units. Thus, the computing device 900 can include discrete components (e.g., a full GPU dedicated to the computing device 900), virtual components (e.g., a portion of a GPU dedicated to the computing device 900), or a combination thereof.

[0085] Although FIG. 9 The various blocks shown in the computing device are not intended to be limiting, and are merely for clarity. For example, in some embodiments, a presentation component 918 such as a display device can be considered an I / O component 914 (e.g., if the display is a touchscreen). As another example, a CPU 906 and / or GPU 908 can include memory (e.g., the memory 904 can represent a storage device in addition to the memory of the GPU 908, CPU 906, and / or other components). In other words, FIG. 9 The computing device is merely illustrative. There can be many variations of the computing device. For example, the computing device can be a workstation, a server, a laptop, a desktop, a tablet, a client device, a mobile device, a hand-held device, a game console, an electronic control unit (ECU), a virtual reality system, and / or other device or system types, as well as combinations thereof. The disclosure is not limited to the components and configurations of the computing device shown in FIG. 9 The computing device is merely illustrative. There can be many variations of the computing device. For example, the computing device can be a workstation, a server, a laptop, a desktop, a tablet, a client device, a mobile device, a hand-held device, a game console, an electronic control unit (ECU), a virtual reality system, and / or other device or system types, as well as combinations thereof. The disclosure is not limited to the components and configurations of the computing device shown in

[0086] The interconnection system 902 can represent one or more links or buses, such as an address bus, a data bus, a control bus, or a combination thereof. The interconnection system 902 can include one or more bus or link types, such as an Industry Standard Architecture (ISA) bus, an Extended Industry Standard Architecture (EISA) bus, a Video Electronics Standards Association (VESA) bus, a Peripheral Component Interconnect (PCI) bus, a Peripheral Component Interconnect Express (PCIe) bus, and / or another type of bus or link. In some embodiments, there are direct connections between components. As an example, the CPU 906 can be directly connected to the memory 904. Further, the CPU 906 can be directly connected to the GPU 908. Where there are direct or point-to-point connections between components, the interconnection system 902 can include a PCIe link to perform the connection. In these examples, a PCI bus need not be included in the computing device 900.

[0087] The memory 904 can include any of a wide variety of computer-readable media. Computer-readable media can be any available media that can be accessed by the computing device 900. Computer-readable media can include both volatile and nonvolatile media, and removable and non-removable media. By way of example, and not limitation, computer-readable media can comprise computer storage media and communication media.

[0088] Computer storage media can include volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, and / or other data types. For example, the memory 904 can store computer readable instructions such as those representing programs and / or program elements, e.g., an operating system. Computer storage media can include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and that can be accessed by the computing device 900. Computer storage media, as used herein, does not include signals per se.

[0089] Computer storage media can include computer-readable instructions, data structures, program modules, and / or other data in a modulated data signal, such as a carrier wave or other transport mechanism, and includes any information delivery media. The term "modulated data signal" can refer to a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, computer storage media can include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media. Combinations of the any of the above should also be included within the scope of computer readable media.

[0090] The CPUs 906 can be configured to execute at least some of the computer-readable instructions in order to control one or more components of the computing device 900 to perform one or more of the methods and / or processes described herein. Each of the CPUs 906 can include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of handling a large number of software threads concurrently. The CPUs 906 can include any type of processors and can include different types of processors depending on the type of computing device 900 being implemented (e.g., processors with fewer cores for mobile devices and processors with more cores for servers). For example, depending on the type of computing device 900, the processors can be Advanced RISC Machines (ARM) processors implemented using Reduced Instruction Set Computing (RISC) or x86 processors implemented using Complex Instruction Set Computing (CISC). The computing device 900 can include one or more CPUs 906 in addition to one or more microprocessors or supplemental co-processors such as math co-processors.

[0091] In addition to or in place of CPU 906, GPU 908 can be configured to execute at least some computer-readable instructions to control one or more components of computing device 900 to perform one or more of the methods and / or processes described herein. One or more of GPUs 908 can be integrated GPUs (e.g., as opposed to one or more of CPUs 906 and / or one or more of GPUs 908 can be discrete GPUs). In embodiments, one or more of GPUs 908 can be a co-processor to one or more of CPUs 906. GPUs 908 can be used by computing device 900 to render graphics (e.g., 3D graphics) or to perform general purpose computing. For example, GPUs 908 can be used for general purpose computing on GPUs (GPGPU). GPUs 908 can include hundreds or thousands of cores capable of processing hundreds or thousands of software threads concurrently. GPUs 908 can generate pixel data for output images in response to rendering commands (e.g., received from CPUs 906 via a host interface). GPUs 908 can include graphics memory, such as display memory, for storing pixel data or any other suitable data (such as GPGPU data). Display memory can be included as part of memory 904. GPUs 908 can include two or more GPUs operating in parallel (e.g., via a link). The link can connect the GPUs directly (e.g., using NVLINK) or can connect the GPUs through a switch (e.g., using NVSwitch). When combined together, each GPU 908 can generate different portions of pixel data or GPGPU data for output or for different outputs (e.g., a first GPU for a first image and a second GPU for a second image). Each GPU can include its own memory or can share memory with other GPUs.

[0092] In addition to or in place of CPU(s) 906 and / or GPU(s) 908, logic unit(s) 920 can be configured to execute at least some of the computer-readable instructions to control one or more components of computing device 900 to perform one or more of the methods and / or processes described herein. In embodiments, CPU(s) 906, GPU(s) 908, and / or logic unit(s) 920 can perform any combination of methods, processes, and / or portions thereof, discretely or jointly. One or more of logic units 920 can be one or more of CPU(s) 906 and / or GPU(s) 908 and / or integrated in one or more of CPU(s) 906 and / or GPU(s) 908, and / or one or more of logic units 920 can be discrete components or otherwise external to CPU(s) 906 and / or GPU(s) 908. In embodiments, one or more of logic units 920 can be a co-processor of one or more of CPU(s) 906 and / or one or more of GPU(s) 908.

[0093] Examples of logic units 920 include one or more processing cores and / or components thereof, such as tensor cores (TCs), tensor processing units (TPUs), pixel vision cores (PVCs), vision processing units (VPUs), graphics processing clusters (GPCs), texture processing clusters (TPCs), streaming multi-processors (SMs), tree traversal units (TTUs), artificial intelligence accelerators (AIAs), deep learning accelerators (DLAs), arithmetic logic units (ALUs), application-specific integrated circuits (ASICs), floating point units (FPUs), input / output (I / O) elements, peripheral component interconnect (PCI) or peripheral component interconnect express (PCIe) elements, and / or the like.

[0094] Communication interface 910 can include one or more receivers, transmitters, and / or transceivers that enable computing device 900 to communicate with other computing devices via electronic communication networks, including wired and / or wireless communication. Communication interface 910 can include components and functionality to enable communication over any of a number of different networks, such as wireless networks (e.g., Wi-Fi, Z-Wave, Bluetooth, Bluetooth LE, ZigBee, etc.), wired networks (e.g., through Ethernet or InfiniBand communication), low power wide area networks (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.

[0095] I / O ports 912 can enable the computing device 900 to logically couple to other devices including I / O components 914, presentation components 918, and / or other components, some of which can be built in to (e.g., integrated in) the computing device 900. Illustrative I / O components 914 include a microphone, mouse, keyboard, joystick, game pad, game controller, satellite dish, scanner, printer, wireless device, etc. The I / O components 914 can provide a natural user interface (NUI) that processes air gestures, voice, or other physiological inputs generated by a user. In some examples, inputs can be transmitted to an appropriate network element for further processing. A NUI can implement any combination of speech recognition, handwriting recognition, facial recognition, biometric recognition, gesture recognition both on screen and adjacent to the screen, air gestures, head and eye tracking, and touch recognition (as described in more detail below) associated with a display of the computing device 900. The computing device 900 can include depth cameras, infrared cameras, RGB cameras, touch screens, and combinations of these, such as a stereoscopic camera system to provide a depth map.

[0096] A power supply 916 can include a hard-wired power supply, a battery power supply, or a combination thereof. The power supply 916 can supply power to the computing device 900 to enable the components of the computing device 900 to operate.

[0097] The presentation components 918 can include a display (e.g., a monitor, a touch screen, a television, a heads-up display (HUD), other display types, or combinations thereof), speakers, and / or other presentation components. The presentation components 918 can receive data from other components (e.g., the GPU 908, the CPU 906, etc.) and output the data (e.g., as a

[0098] Example data center

[0099] FIG. 10 An example data center 1000 that can be used in at least one embodiment of the present disclosure is illustrated. The data center 1000 includes a data center infrastructure layer 1010, a framework layer 1020, a software layer 1030, and an application layer 1040.

[0100] As FIG. 10As shown, the data center infrastructure layer 1010 can include a resource orchestrator 1022, grouped computing resources 1014, and node computing resources ("node C.R.s") 1016(1)-1016(N), where "N" represents a positive integer (which can be a different integer "N" than the integer used in the other figures). In at least one embodiment, the node C.R.s 1016(1)-1016(N) can include, without limitation, any number of central processing units ("CPUs") or other processors (including accelerators, field programmable gate arrays (FPGAs), graphics processors or graphics processing units (GPUs), etc.), memory devices (e.g., dynamic random access memory), storage devices (e.g., solid state or disk drives), network input / output ("NW I / O") devices, network switches, virtual machines ("VMs"), power modules, and / or cooling modules, etc. In some embodiments, one or more of the node C.R.s 1016(1)-1016(N) can correspond to a server having one or more of the above-described computing resources. Moreover, in some embodiments, the node C.R.s 1016(1)-1016(N) can include one or more virtual components, such as a vGPU, a vCPU, etc., and / or one or more of the node C.R.s 1016(1)-1016(N) can correspond to a virtual machine (VM).

[0101] In at least one embodiment, the grouped computing resources 1014 can include separate groupings of node C.R.s 1016 housed within one or more racks (not shown), or housed within a number of racks (also not shown) within various geographic locations. The separate groupings of node C.R.s 1016 within the grouped computing resources 1014 can include groupings of computing, network, memory, or storage resources that can be configured or allocated to support one or more workloads. In at least one embodiment, several node C.R.s 1016 including CPUs, GPUs, and / or processors can be grouped within one or more racks to provide computing resources to support one or more workloads. The one or more racks can also include any number and combination of power modules, cooling modules, and / or network switches.

[0102] The resource orchestrator 1022 can configure or otherwise control the one or more node C.R.s 1016(1)-1016(N) and / or the grouped computing resources 1014. In at least one embodiment, the resource orchestrator 1022 can include a software design infrastructure ("SDI") management entity for the data center 1000. In at least one embodiment, the resource orchestrator 1022 can comprise hardware, software, or some combination thereof.

[0103] In at least one embodiment, as FIG. 10As shown, the framework layer 1020 includes a job scheduler 1032, a configuration manager 1034, a resource manager 1036, and a distributed file system 1038. In at least one embodiment, the framework layer 1020 can include a framework that supports the software 1032 of the software layer 1030 and / or one or more applications 1042 of the application layer 1040. The software 1032 or the applications 1042 can include, respectively, web-based services software or applications, such as services or applications provided by Amazon Web Services, Google Cloud, and Microsoft Azure. The framework layer 1020 can be, but is not limited to, a type of free and open-source software web application framework such as Apache Spark™ (hereinafter “Spark”) that can utilize the distributed file system 1038 for large-scale data processing (e.g., “big data”). In at least one embodiment, the job scheduler 1032 can include a Spark driver to facilitate scheduling workloads supported by various layers of the data center 1000. The configuration manager 1034 can be capable of configuring different layers, such as the software layer 1030 and the framework layer 1020 including Spark and the distributed file system 1038 for supporting large-scale data processing. The resource manager 1036 can manage clustered or grouped computing resources mapped to or allocated for supporting the distributed file system 1038 and the job scheduler 1032. In at least one embodiment, the clustered or grouped computing resources can include the grouped computing resources 1014 on the data center infrastructure layer 1010. The resource manager 1036 can coordinate with the resource orchestrator 1012 to manage these mapped or allocated computing resources.

[0104] In at least one embodiment, the software 1032 included in the software layer 1030 can include software used by at least a portion of the node C.R.s 1016(1)-1016(N), the grouped computing resources 1014, and / or the distributed file system 1038 of the framework layer 1020. One or more types of software can include, but are not limited to, Internet web page search software, e-mail virus scanning software, database software, and streaming video content software.

[0105] In at least one embodiment, one or more applications 1042 included in application layer 1040 can include one or more types of applications used by at least portions of node C.R. 1016(1)-1016(N), grouped computing resources 1014, and / or distributed file system 1038 of framework layer 1020. One or more types of applications can include, but are not limited to, any number and type of genomics applications, cognitive computing, applications, and machine learning applications including training or inferencing software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), or other machine learning applications used in conjunction with one or more embodiments.

[0106] In at least one embodiment, any of configuration manager 1034, resource manager 1036, and resource orchestrator 1012 can implement any number and type of self-modifying actions based on any amount and type of data acquired in any technically feasible manner. Self-modifying actions can relieve data center operators of data center 1000 from making possibly poor configuration decisions and can avoid underutilization and / or poorly performing portions of a data center.

[0107] Data center 1000 can include tools, services, software, or other resources to train one or more machine learning models or use one or more machine learning models to predict or infer information in accordance with one or more embodiments described herein. For example, a machine learning model can be trained by computing weight parameters according to a neural network architecture using software and / or computing resources described above with respect to data center 1000. In at least one embodiment, using trained or deployed machine learning models corresponding to one or more neural networks can be used to infer or predict information using resources described above with respect to data center 1000 using weight parameters computed by one or more training techniques such as, but not limited to, those described herein.

[0108] In at least one embodiment, data center 100 can use CPUs, application specific integrated circuits (ASICs), GPUs, FPGAs, or other hardware (or virtual computing resources corresponding thereto) to perform training and / or inference using resources described above. Furthermore, one or more software and / or hardware resources described above can be configured as a service to allow users to train or perform information inference such as image recognition, speech recognition, or other artificial intelligence services.

[0109] Example Network Environment

[0110] Network environments suitable for use in implementing embodiments of the present disclosure can include one or more client devices, servers, network-attached storage (NAS), other backend devices, and / or other device types. The client devices, servers, and / or other device types (e.g., each device) can be implemented on one or more instances of the computing device(s) 900— e.g., each device can include similar components, features, and / or functionality of the computing device(s) 900. Further, where a backend device (e.g., a server, NAS, etc.) is implemented, the backend device can be included as part of a data center 1000, an example of which is described in greater detail herein with respect to FIG. 1. FIG. 9 FIG. 10

[0111] Components of the network environment can communicate with each other via network(s), which can be wired, wireless, or both. The network(s) can include multiple networks or one of multiple networks. For example, the network(s) can include one or more wide area networks (WANs), one or more local area networks (LANs), one or more public networks such as the Internet, and / or the public switched telephone network (PSTN) and / or one or more private networks. Where the network(s) include a wireless telecommunication network, components such as base stations, communication towers, or even access points (among other components) can provide wireless connectivity.

[0112] Compatible network environments can include one or more peer-to-peer network environments (in which case servers can not be included in the network environment) and one or more client-server network environments (in which case one or more servers can be included in the network environment). In a peer-to-peer network environment, functionality described herein with respect to servers can be implemented on any number of client devices.

[0113] ​​In at least one embodiment, a network environment can include one or more cloud-based network environments, distributed computing environments, combinations thereof, and the like. A cloud-based network environment can include a framework layer, a work scheduler, a resource manager, and a distributed file system implemented on one or more servers, which can include one or more core network servers and / or edge servers. The framework layer can include a framework that supports a software layer and / or one or more applications of an application layer. The software or applications can include network-based service software or applications, respectively. In embodiments, one or more client devices can use the network-based service software or applications (e.g., by accessing the service software and / or applications via one or more application programming interfaces (APIs)). The framework layer can be, without limitation, a type of free and open-source software web application framework, such as can use a distributed file system for large-scale data processing (e.g., “big data”).

[0114] A cloud-based network environment can provide cloud computing and / or cloud storage that performs any combination of the computing and / or data storage functions described herein (or one or more portions thereof). Any of these different functions can be distributed across multiple locations from central or core servers (e.g., one or more data centers that can be distributed across states, regions, countries, the Earth, and the like). Core servers can designate at least a portion of a function to an edge server if a connection to a user (e.g., a client device) is relatively close to the edge server. A cloud-based network environment can be private (e.g., limited to a single organization), can be public (e.g., available to many organizations), and / or combinations thereof (e.g., a hybrid cloud environment).

[0115] A client device can include at least some components, features, and functionality of the example computing device(s) 900 described herein. FIG. 9 By way of example, and not limitation, a client device can be implemented as a personal computer (PC), a laptop computer, a mobile device, a smart phone, a tablet computer, a smart watch, a wearable computer, a personal digital assistant (PDA), an MP3 player, a virtual reality headset, a global positioning system (GPS) or device, a video player, a camera, a surveillance device or system, a vehicle, a ship, a spacecraft, a virtual machine, a drone, a robot, a hand-held communication device, a hospital device, a gaming device or system, an entertainment system, a vehicle computer system, an embedded system controller, a remote control, an appliance, a consumer electronic device, a workstation, an edge device, any combination of these depicted devices, or any other suitable device.

[0116] Example autonomous vehicle

[0117] FIG. 11AAn illustration of an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. Autonomous vehicle 1100 (alternatively referred to herein as “vehicle 1100”) can include, but is not limited to, a passenger vehicle such as a car, truck, bus, ambulance, shuttle, electric or motorized bicycle, motorcycle, fire truck, police car, ambulance, boat, construction vehicle, underwater vehicle, drone, and / or another type of vehicle (e.g., unmanned and / or capable of accommodating one or more passengers). Autonomous vehicles are often described in terms of levels of automation as defined by a department of the United States Department of Transportation, the National Highway Traffic Safety Administration (NHTSA), and the Society of Automotive Engineers (SAE) “Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles” (Standard No. J3016-201806 published June 15, 2018, Standard No. J3016-201609 published September 30, 2016, and previous and future versions of this standard). Vehicle 1100 can be capable of implementing functionality that complies with one or more of Levels 3-5 of autonomous driving. For example, depending on the embodiment, vehicle 1100 can be capable of conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5).

[0118] Vehicle 1100 can include components such as a chassis, a body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components of a vehicle. Vehicle 1100 can include a propulsion system 1150 such as an internal combustion engine, a hybrid power plant, a fully electric motor, and / or another type of propulsion system. Propulsion system 1150 can be connected to a drivetrain of vehicle 1100 that can include a transmission in order to effect propulsion of vehicle 1100. Propulsion system 1150 can be controlled in response to receiving a signal from a throttle / accelerator 1152.

[0119] A steering system 1154, which can include a steering wheel, can be used to steer vehicle 1100 (e.g., along a desired path or route) while propulsion system 1150 is operating (e.g., while the vehicle is in motion). Steering system 1154 can receive a signal from a steering actuator 1156. For full automation (Level 5) functionality, a steering wheel can be optional.

[0120] A braking sensor system 1146 can be used to operate vehicle brakes in response to receiving a signal from a braking actuator 1148 and / or a braking sensor.

[0121] may include one or more system on chips (SoCs) 1104 (FIG. 11C ) and / or one or more controllers 1136 of one or more GPUs can provide signals (e.g., representative of commands) to one or more components and / or systems of the vehicle 1100. For example, the one or more controllers can send signals to operate vehicle brakes via one or more brake actuators 1148, to operate a steering system 1154 via one or more steering actuators 1156, to operate a propulsion system 1150 via one or more throttle / accelerator 1152. The one or more controllers 1136 can include one or more on-board (e.g., integrated) computing devices (e.g., supercomputers) that process sensor signals and output operational commands (e.g., signals representative of commands) to enable autonomous driving and / or to assist a human driver in driving the vehicle 1100. The one or more controllers 1136 can include a first controller 1136 for autonomous driving functions, a second controller 1136 for functional safety functions, a third controller 1136 for artificial intelligence functions (e.g., computer vision), a fourth controller 1136 for infotainment functions, a fifth controller 1136 for redundancy in emergency situations, and / or other controllers. In some examples, a single controller 1136 can handle two or more of the above functions, two or more controllers 1136 can handle a single function, and / or any combination thereof.

[0122] The one or more controllers 1136 can provide signals for controlling one or more components and / or systems of the vehicle 1100 in response to sensor data (e.g., sensor inputs) received from one or more sensors. The sensor data can be received from, for example and without limitation, a global navigation satellite system sensor 1158 (e.g., a global positioning system sensor), a RADAR sensor 1160, an ultrasonic sensor 1162, a LIDAR sensor 1164, an inertial measurement unit (IMU) sensor 1166 (e.g., an accelerometer, a gyroscope, a magnetic compass, a magnetometer, etc.), a microphone 1196, a stereo camera 1168, a wide-angle camera 1170 (e.g., a fisheye camera), an infrared camera 1172, a surround camera 1174 (e.g., a 360-degree camera), a long and / or medium range camera 1198, a speed sensor 1144 (e.g., to measure a speed of the vehicle 1100), a vibration sensor 1142, a steering sensor 1140, a brake sensor (e.g., as part of a brake sensor system 1146), and / or other sensor types.

[0123] One or more of the controllers 1136 can receive input (e.g., represented by input data) from the instrument cluster 1132 of the vehicle 1100 and provide output (e.g., represented by output data, display data, etc.) via the human-machine interface (HMI) display 1134, audible annunciators, speakers, and / or via other components of the vehicle 1100. These outputs can include information such as vehicle speed, velocity, time, map data (e.g., HD map 1122), location data (e.g., location of the vehicle 1100, e.g., on a map), direction, location of other vehicles (e.g., an occupancy grid), information about objects and object states as perceived by the controllers 1136, and so on. For example, the HMI display 1134 can display information about the presence of one or more objects (e.g., street signs, warning signs, traffic light changes, etc.) and / or information about driving maneuvers that the vehicle has made, is making, or will make (e.g., change lanes now, exit 34B in two miles, etc.). FIG. 11C

[0124] The vehicle 1100 further includes a network interface 1124 that can communicate over one or more networks using one or more wireless antennas 1126 and / or modems. For example, the network interface 1124 can be capable of communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, etc. The one or more wireless antennas 1126 can also enable communication between objects (e.g., vehicles, mobile devices, etc.) in the environment using one or more local area networks such as Bluetooth, Bluetooth LE, Z-Wave, ZigBee, etc. and / or one or more low power wide area networks (LPWANs) such as LoRaWAN, SigFox, etc.

[0125] FIG. 11B An example autonomous vehicle 1100 for use in accordance with some embodiments of the present disclosure. FIG. 11A An example camera position and field of view for the example autonomous vehicle 1100 for use in accordance with some embodiments of the present disclosure.

[0126] ​Camera types for the cameras can include, but are not limited to, digital cameras that can be adapted for use with components and / or systems of the vehicle 1100. The cameras can operate at Automotive Safety Integrity Level (ASIL) B and / or at another ASIL. The camera types can have any image capture rate, such as 60 frames per second (fps), 120 fps, 240 fps, etc., depending on the embodiment. The cameras can be capable of using a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In some examples, a color filter array can include a red-white-white-white (RCCC) color filter array, a red-white-white-blue (RCCB) color filter array, a red-blue-green-white (RBGC) color filter array, a Foveon X3 color filter array, a Bayer sensor (RGGB) color filter array, a monochrome sensor color filter array, and / or another type of color filter array. In some embodiments, clear pixel cameras, such as cameras with a CCC, RCCB, and / or RBGC color filter array, can be used in efforts to improve light sensitivity.

[0127] In some examples, one or more of the cameras can be used to perform advanced driver assistance system (ADAS) functions (e.g., as part of a redundant or fail-safe design). For example, a multi-functional mono camera can be installed to provide functions including lane departure warning, traffic sign assist, and intelligent headlamp control. One or more of the cameras (e.g., all of the cameras) can simultaneously record and provide image data (e.g., video).

[0128] One or more of the cameras can be installed in mounting assemblies, such as custom designed (3-D printed) assemblies, in order to cut off stray light and reflections from within the car that can interfere with the image data capture capabilities of the cameras (e.g., reflections from the dashboard reflected in the windshield mirror). With regard to wing mirror mounting assemblies, the wing mirror assemblies can be custom 3-D printed such that the camera mounting plates match the shape of the wing mirrors. In some examples, one or more cameras can be integrated into the wing mirrors. For side view cameras, one or more cameras can also be integrated into the four pillars of each corner of the cab.

[0129] A camera with a field of view that includes the environment in front of the vehicle 1100 (e.g., a front-facing camera) can be used for surround view to help identify forward paths and obstacles, and, with the assistance of one or more controllers 1136 and / or control SoCs, to provide information crucial for generating an occupancy grid and / or determining a preferred vehicle path. The front-facing camera can be used to perform many of the same ADAS functions as LiDAR, including emergency braking, pedestrian detection, and collision avoidance. The front-facing camera can also be used in ADAS functions and systems, including Lane Departure Warning (“LDW”), Autonomous Cruise Control (“ACC”), and / or other functions such as traffic sign recognition.

[0130] A variety of cameras can be used in front-facing configurations, including, for example, monocular camera platforms including CMOS (Complementary Metal-Oxide-Semiconductor) color imagers. Another example could be a wide-angle camera 1170, which can be used to perceive objects entering the field of view from the periphery (e.g., pedestrians, traffic at intersections, or bicycles). Although FIG. 11B The image shows only one wide-angle camera, but any number of wide-angle cameras 1170 can be present on vehicle 1100. Furthermore, a remote camera 1198 (e.g., a pair of long-view stereo cameras) can be used for depth-based object detection, especially for objects for which a neural network has not yet been trained. The remote camera 1198 can also be used for object detection and classification, as well as basic object tracking.

[0131] One or more stereo cameras 1168 may also be included in a front-mounted configuration. The stereo camera 1168 may include an integrated control unit comprising a scalable processing unit that can provide a multi-core microprocessor and programmable logic (FPGA) with an integrated CAN or Ethernet interface on a single chip. Such a unit can be used to generate a 3D map of the vehicle environment, including distance estimates for all points in the image. Alternative stereo cameras 1168 may include a compact stereo vision sensor that may include two camera lenses (one on each side) and an image processing chip capable of measuring the distance from the vehicle to a target object and using the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. Other types of stereo cameras 1168 may be used in addition to those described herein, or alternatively.

[0132] A camera (e.g., a side-view camera) having a field of view including the side of the vehicle 1100 can be used for surround view, providing information for creating and updating occupancy grids and generating side-impact collision warnings. For example, a surround camera 1174 (e.g., ...) FIG. 11BFour surround cameras 1174) can be placed on the vehicle 1100. The surround cameras 1174 can include wide-view cameras 1170, fisheye cameras, 360-degree cameras, and / or the like. In one example, four fisheye cameras can be placed on the front, back, and sides of the vehicle. In an alternative arrangement, the vehicle can use three surround cameras 1174 (e.g., left, right, and back) and can utilize one or more other cameras (e.g., a forward-facing camera) as the fourth surround view camera.

[0133] Cameras with a field of view that includes an environmental portion of the back of the vehicle 1100 (e.g., rearview cameras) can be used to assist with parking, surround view, back collision warnings, and creating and updating the occupancy grid. A wide variety of cameras can be used, including but not limited to cameras that are also suitable as front-facing cameras (e.g., long and / or mid-range cameras 1198, stereo cameras 1168, infrared cameras 1172, etc.) as described herein.

[0134] FIG. 11C FIG. 1 illustrates an example autonomous vehicle 1100 for use in accordance with some embodiments of the present disclosure. FIG. 11A FIG. 1 illustrates an example autonomous vehicle 1100 for use in accordance with some embodiments of the present disclosure.

[0135] FIG. 11C Each of the components, features, and systems of the vehicle 1100 in FIG. 1 are shown connected via a bus 1102. The bus 1102 can include a controller area network (CAN) data interface (alternatively referred to herein as a "CAN bus"). The CAN can be a network within the vehicle 1100 that is used to assist in controlling various features and functions of the vehicle 1100, such as the actuation of brakes, acceleration, braking, steering, windshield wipers, and the like. The CAN bus can be configured to have tens or even hundreds of nodes, each with its own unique identifier (e.g., CAN ID). The CAN bus can be read to find steering wheel angle, ground speed, engine revolutions per minute (RPM), button positions, and / or other vehicle status indicators. The CAN bus can be ASIL B compliant.

[0136] Although bus 1102 is described herein as a CAN bus, this is not intended to be limiting. For example, FlexRay and / or Ethernet can be used in addition to or instead of a CAN bus. Further, although bus 1102 is represented with a single line, this is not intended to be limiting. For example, there can be any number of buses 1102, which can include one or more CAN buses, one or more FlexRay buses, one or more Ethernet buses, and / or one or more other types of buses that use different protocols. In some examples, two or more buses 1102 can be used to perform different functions, and / or can be used for redundancy. For example, a first bus 1102 can be used for collision avoidance functions, and a second bus 1102 can be used for drive control. In any example, each bus 1102 can communicate with any component of vehicle 1100, and two or more buses 1102 can communicate with the same components. In some examples, each SoC 1104, each controller 1136, and / or each computer within the vehicle can have access to the same input data (e.g., input from sensors of vehicle 1100), and can be connected to a common bus, such as a CAN bus.

[0137] Vehicle 1100 can include one or more controllers 1136, such as those described herein with respect to FIG. 11A controllers. Controllers 1136 can be used for a wide variety of functions. Controllers 1136 can be coupled to any other distinct component and system of vehicle 1100, and can be used for control of vehicle 1100, artificial intelligence of vehicle 1100, infotainment for vehicle 1100, and / or the like.

[0138] Vehicle 1100 can include one or more system on chips (SoCs) 1104. SoCs 1104 can include CPUs 1106, GPUs 1108, processors 1110, caches 1112, accelerators 1114, data stores 1116, and / or other components and features not shown. In a wide variety of platforms and systems, SoCs 1104 can be used to control vehicle 1100. For example, one or more SoCs 1104 can be used in a system (e.g., a system of vehicle 1100) in conjunction with HD map 1122, which can obtain map refreshes and / or updates from one or more servers (e.g., one or more servers 1178) via network interface 1124. FIG. 11D

[0139] ​CPU 1106 can include a CPU cluster or CPU complex (alternatively referred to herein as a "CCPLEX"). CPU 1106 can include multiple cores and / or L2 caches. For example, in some embodiments, CPU 1106 can include eight cores in a coherent multi-processor configuration. In some embodiments, CPU 1106 can include four dual-core clusters, with each cluster having a dedicated L2 cache (e.g., a 2 MB L2 cache). CPU 1106 (e.g., the CCPLEX) can be configured to support simultaneous cluster operation, such that any combination of clusters of CPU 1106 can be active at any given time.

[0140] CPU 1106 can implement power management capabilities including one or more of the following features: individual hardware blocks can be automatically clock-gated when idle to save dynamic power; each core clock can be gated when the core is not actively executing instructions due to execution of WFI / WFE instructions; each core can be independently power-gated; each core cluster can be independently clock-gated when all cores are clock-gated or power-gated; and / or each core cluster can be independently power-gated when all cores are power-gated. CPU 1106 can further implement enhanced algorithms for managing power states, with specified allowed power states and desired wake-up times, and the hardware / microcode determines the optimal power state for the cores, clusters, and CCPLEX to enter. The processing cores can support a simplified power state entry sequence in software, with the work offloaded to microcode.

[0141] GPU 1108 can include an integrated GPU (alternatively referred to herein as an "iGPU"). GPU 1108 can be programmable and efficient for parallel workloads. In some examples, GPU 1108 can use an enhanced tensor instruction set. GPU 1108 can include one or more streaming microprocessors, where each streaming microprocessor can include an Ll cache (e.g., an Ll cache having at least 96 KB of storage capacity), and two or more of the streaming microprocessors can share an L2 cache (e.g., an L2 cache having 512 KB of storage capacity). In some embodiments, GPU 1108 can include at least eight streaming microprocessors. GPU 1108 can use a compute application programming interface (API). In addition, GPU 1108 can use one or more parallel computing platforms and / or programming models (e.g., NVIDIA's CUDA).

[0142] In the case of automotive and embedded uses, the GPU 1108 can be power-optimized for best performance. For example, the GPU 1108 can be fabricated on a fin field-effect transistor (FinFET) for lower power consumption. However, this is not intended to be limiting, and the GPU 1108 can be fabricated using other semiconductor manufacturing processes. Each streaming microprocessor can incorporate several mixed-precision processing cores divided into multiple blocks. For example, and without limitation, 64 PF32 cores and 32 PF64 cores can be divided into four processing blocks. In such an example, each processing block can be allocated 16 FP32 cores, 8 FP64 cores, 16 INT32 cores, two mixed-precision NVIDIA Tensor Cores for deep learning matrix arithmetic, an L0 instruction cache, a thread warp scheduler, a dispatch unit, and / or a 64 KB register file. Further, the streaming microprocessor can include independent parallel integer and floating point data paths to exploit the mix of computation and address computation for efficient execution of workloads. The streaming microprocessor can include independent thread scheduling capabilities to allow for finer-grain synchronization and cooperation between parallel threads. The streaming microprocessor can include a combined LI data cache and shared memory unit to improve performance while simplifying programming.

[0143] The GPU 1108 can include a high bandwidth memory (HBM) and / or a 16 GB HBM2 memory subsystem that provides approximately 900 GB / s of peak memory bandwidth in some examples. In some examples, in addition to or alternatively from HBM memory, synchronous graphics random access memory (SGRAM) can be used, such as fifth generation graphics double data rate synchronous random access memory (GDDR5).

[0144] The GPU 1108 can include a unified memory technology that includes access counters to allow memory pages to be migrated more precisely to the processors that access them most frequently, improving efficiency of memory ranges shared between processors. In some examples, address translation services (ATS) support can be used to allow the GPU 1108 to access CPU 1106 page tables directly. In such examples, when the GPU 1108 memory management unit (MMU) experiences a miss, an address translation request can be transmitted to the CPU 1106. In response, the CPU 1106 can look up the virtual-to-physical mapping for the address in its page tables and transmit the translation back to the GPU 1108. In this way, the unified memory technology can allow a single unified virtual address space for memory of both the CPU 1106 and the GPU 1108, simplifying GPU 1108 programming and porting applications to the GPU 1108.

[0145] In addition, GPU 1108 can include an access counter that can track how often GPU 1108 accesses memory of other processors. The access counter can help ensure that memory pages are moved to the physical memory of the processor that most frequently accesses those pages.

[0146] SoC 1104 can include any number of caches 1112, including those described herein. For example, caches 1112 can include an L3 cache that is available to both CPU 1106 and GPU 1108 (e.g., connected to both CPU 1106 and GPU 1108). Caches 1112 can include a write-back cache that can track the state of a line, for example, by using a cache coherency protocol (e.g., MEI, MESI, MSI, etc.). Depending on the embodiment, the L3 cache can include 4MB or more, although smaller cache sizes can also be used.

[0147] SoC 1104 can include an arithmetic logic unit (ALU) that can be used to perform processing with respect to any of a variety of tasks or operations of vehicle 1100, such as processing a DNN. In addition, SoC 1104 can include a floating point unit (FPU) - or other mathematical co-processor or digital co-processor type - for performing mathematical operations within the system. For example, SoC 104 can include one or more FPUs integrated as execution units within CPU 1106 and / or GPU 1108.

[0148] SoC 1104 can include one or more accelerators 1114 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, SoC 1104 can include a hardware acceleration cluster that can include optimized hardware accelerators and / or a large on-chip memory. This large on-chip memory (e.g., 4MB SRAM) can enable the hardware acceleration cluster to accelerate neural networks and other computations. The hardware acceleration cluster can be used to supplement GPU 1108 and offload some of the tasks of GPU 1108 (e.g., freeing up more cycles of GPU 1108 for performing other tasks). As one example, accelerators 1114 can be used for targeted workloads (e.g., perception, convolutional neural networks (CNNs), etc.) that are stable enough to accelerate easily. As used herein, the term “CNN” can include all types of CNNs, including region-based or region convolutional neural networks (RCNNs) and fast RCNNs (e.g., for object detection).

[0149] The accelerators 1114 (e.g., hardware acceleration cluster) can include a deep learning accelerator (DLA). The DLA can include one or more tensor processing units (TPUs) that can be configured to provide an additional 100 billion operations per second for deep learning applications and inferencing. The TPU can be an accelerator that is configured to perform image processing functions (e.g., for CNNs, RCNNs, etc.) and is optimized for performing image processing functions. The DLA can be further optimized for a specific set of neural network types and floating point operations and inferencing. The design of the DLA can provide higher performance per mm than general purpose GPUs and far exceeds the performance of CPUs. The TPU can perform several functions including single instance convolution functions, support for INT8, INT16, and FP16 data types for both features and weights, for example, and post-processor functions.

[0150] The DLA can perform neural networks, especially CNNs, on processed or unprocessed data for any of a wide variety of functions, such as and not by way of limitation: CNNs for object recognition and detection using data from camera sensors; CNNs for distance estimation using data from camera sensors; CNNs for emergency vehicle detection and identification and detection using data from microphones; CNNs for facial recognition and vehicle owner identification using data from camera sensors; and / or CNNs for safety and / or safety related events.

[0151] The DLA can perform any of the functions of the GPU 1108, and by using an inferencing accelerator, the designer can target the DLA or the GPU 1108 for any function, for example. For example, the designer can focus the processing and floating point operations of the CNNs on the DLA and leave other functions to the GPU 1108 and / or other accelerators 1114.

[0152] The accelerators 1114 (e.g., hardware acceleration cluster) can include a programmable vision accelerator (PVA), which can be alternatively referred to herein as a computer vision accelerator. The PVA can be designed and configured to accelerate computer vision algorithms for advanced driver assistance systems (ADAS), autonomous driving, and / or augmented reality (AR) and / or virtual reality (VR) applications. The PVA can provide a balance between performance and flexibility. For example, each PVA can include any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors, for example, and not by way of limitation.

[0153] The RISC cores can interact with image sensors (e.g., image sensors of any of the cameras described herein), image signal processors, and / or the like. Each of these RISC cores can include any number of memories. Depending on the embodiment, the RISC cores can use any of several protocols. In some examples, the RISC cores can execute a real-time operating system (RTOS). The RISC cores can be implemented using one or more integrated circuit devices, application specific integrated circuits (ASICs), and / or memory devices. For example, the RISC cores can include instruction caches and / or tightly coupled RAM.

[0154] The DMA can enable components of the PVA to access system memory independently of the CPU 1106. The DMA can support any number of features to provide optimization to the PVA, including but not limited to supporting multi-dimensional addressing and / or circular addressing. In some examples, the DMA can support addressing up to six or more dimensions, which can include block width, block height, block depth, horizontal block step, vertical block step, and / or depth step.

[0155] The vector processor can be a programmable processor that can be designed to efficiently and flexibly execute programming for computer vision algorithms and provide signal processing capabilities. In some examples, the PVA can include a PVA core and two vector processing subsystem partitions. The PVA core can include a processor subsystem, one or more DMA engines (e.g., two DMA engines), and / or other peripherals. The vector processing subsystems can operate as the main processing engines of the PVA and can include a vector processing unit (VPU), an instruction cache, and / or a vector memory (e.g., VMEM). The VPU core can include a digital signal processor such as, for example, a single instruction multiple data (SIMD), very long instruction word (VLIW) digital signal processor. The combination of SIMD and VLIW can enhance throughput and speed.

[0156] Each of the vector processors can include an instruction cache and can be coupled to a dedicated memory. As a result, in some examples, each of the vector processors can be configured to execute independently of the other vector processors. In other examples, the vector processors included in a particular PVA can be configured to employ data parallelization. For example, in some embodiments, multiple vector processors included in a single PVA can execute the same computer vision algorithm, but on different regions of an image. In other examples, the vector processors included in a particular PVA can execute different computer vision algorithms on the same image simultaneously, or even different algorithms on a sequence of images or portions of an image. Any number of PVAs can be included in the hardware acceleration cluster, and any number of vector processors can be included in each of the PVAs, among other things. Furthermore, the PVAs can include additional error-correcting code (ECC) memory to enhance overall system security.

[0157] The accelerator 1114 (e.g., hardware acceleration cluster) can include an on-chip computer vision network and SRAM to provide high bandwidth, low latency SRAM for the accelerator 1114. In some examples, the on-chip memory can include at least 4 MB of SRAM composed of, for example and without limitation, eight field-programmable memory blocks, which can be accessed by both the PVA and the DLA. Each pair of memory blocks can include an advanced peripheral bus (APB) interface, configuration circuitry, a controller, and a multiplexer. Any type of memory can be used. The PVA and the DLA can access the memory via a backbone that provides high-speed memory access to the PVA and the DLA. The backbone can include an on-chip computer vision network that interconnects the PVA and the DLA to the memory, for example using an APB.

[0158] The on-chip computer vision network can include an interface that determines that both the PVA and the DLA provide ready and valid signals before transmitting any control signals / addresses / data. Such an interface can provide separate phases and separate channels for transmitting control signals / addresses / data, as well as burst communications for continuous data transmission. This type of interface can comply with ISO 26262 or IEC 61508 standards, but other standards and protocols can also be used.

[0159] In some examples, the SoC 1104 can include a real-time ray tracing hardware accelerator, such as described in U.S. Patent Application No. 16 / 101,232, filed August 10, 2018. The real-time ray tracing hardware accelerator can be used to quickly and efficiently determine locations and extents of objects (e.g., within a world model) in order to generate real-time visualizations simulations for RADAR signal interpretation, for sound propagation synthesis and / or analysis, for SONAR system simulation, for general wave propagation simulation, for comparison to LIDAR data for purposes of localization and / or other functionality, and / or for other uses. In some embodiments, one or more tree traversal units (TTUs) can be used to perform one or more ray tracing related operations.

[0160] The accelerator 1114 (e.g., a hardware accelerator cluster) has a wide range of uses for autonomous driving. The PVA can be a programmable vision accelerator that can be used for key processing stages in ADAS and autonomous vehicles. The capabilities of the PVA are a good match for algorithm domains that require predictable processing, low power, and low latency. In other words, the PVA performs well on semi-dense or dense regular computations, and even on small data sets that require predictable runtimes with low latency and low power. Thus, in the context of a platform for autonomous vehicles, the PVA is designed to run classical computer vision algorithms because they are effective at object detection and integer math operations.

[0161] For example, according to one embodiment of the technology, the PVA is used to perform computer stereo vision. In some examples, a semi-global matching based algorithm can be used, although this is not intended to be limiting. Many applications for level 3-5 autonomous driving require instant motion estimation / stereo matching (e.g., structure from motion, pedestrian recognition, lane detection, etc.). The PVA can perform computer stereo vision functions on input from two monocular cameras.

[0162] In some examples, the PVA can be used to perform dense optical flow. Raw RADAR data is processed according to a process (e.g., using a 4D fast Fourier transform) to provide processed RADAR. In other examples, the PVA is used for time-of-flight depth processing, such as by processing raw time-of-flight data to provide processed time-of-flight data.

[0163] The DLA can be used to run any type of network to enhance control and driving safety, including, for example, a neural network that outputs a confidence metric for each object detection. Such a confidence value can be interpreted as a probability, or as providing a relative "weight" for each detection compared to other detections. The confidence value enables the system to make further decisions about which detections should be considered true positive detections rather than false positive detections. For example, the system can set a threshold for confidence, and only consider detections that exceed the threshold as true positive detections. In an automatic emergency braking (AEB) system, false positive detections would cause the vehicle to automatically perform an emergency brake, which is obviously undesirable. Thus, only the most confident detections should be considered a trigger for AEB. The DLA can run a neural network for regression of a confidence value. The neural network can take as its input at least some subset of parameters, such as bounding box dimensions, a ground plane estimate obtained (e.g., from another subsystem), inertial measurement unit (IMU) sensor 1166 outputs related to vehicle 1100 orientation, distance, 3D position estimates of objects obtained from the neural network and / or other sensors (e.g., LIDAR sensor 1164 or RADAR sensor 1160), etc.

[0164] SoC 1104 can include one or more data stores 1116 (e.g., memory). Data stores 1116 can be on-chip memory of SoC 1104, which can store neural networks to be executed on the GPU and / or DLA. In some examples, for redundancy and safety, data stores 1116 can be large enough in capacity to store multiple instances of a neural network. Data stores 1112 can include L2 or L3 cache 1112. References to data stores 1116 can include references to memory associated with PVA, DLA, and / or other accelerators 1114 as described herein.

[0165] SoC 1104 can include one or more processors 1110 (e.g., embedded processors). The processors 1110 can include a boot and power management processor, which can be a specialized processor and subsystem for handling boot power and management functions, as well as security implementation. The boot and power management processor can be part of the SoC 1104 boot sequence and can provide runtime power management services. The boot power and management processor can provide clock and voltage programming, auxiliary system low power state transitions, SoC 1104 thermal and temperature sensor management, and / or SoC 1104 power state management. Each temperature sensor can be implemented as a ring oscillator whose output frequency is proportional to temperature, and the SoC 1104 can use the ring oscillator to detect the temperature of the CPU 1106, GPU 1108, and / or accelerator 1114. If it is determined that the temperature exceeds a threshold, the boot and power management processor can enter a temperature fault routine and place the SoC 1104 in a lower power state and / or place the vehicle 1100 in a driver safe park mode (e.g., safely park the vehicle 1100).

[0166] The processors 1110 can further include a set of embedded processors that can be used as an audio processing engine. The audio processing engine can be an audio subsystem that allows for full hardware support for multi-channel audio over multiple interfaces, as well as a range of extensive and flexible audio I / O interfaces. In some examples, the audio processing engine is a specialized processor core with a digital signal processor with dedicated RAM.

[0167] The processors 1110 can further include an always-on processor engine, which can provide the necessary hardware features to support low-power sensor management and wake-up use cases. The always-on processor engine can include a processor core, tightly coupled RAM, supporting peripherals (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0168] The processors 1110 can further include a security cluster engine, which includes a specialized processor subsystem that handles security management for automotive applications. The security cluster engine can include two or more processor cores, tightly coupled RAM, supporting peripherals (e.g., timers, interrupt controllers, etc.), and / or routing logic. In a secure mode, the two or more cores can operate in a lockstep mode and act as a single core with comparison logic that detects any differences between their operations.

[0169] The processors 1110 can further include a real-time camera engine, which can include a specialized processor subsystem for handling real-time camera management.

[0170] The processor 1110 can further include a high dynamic range signal processor, which can include an image signal processor, which is a hardware engine that is part of the camera processing pipeline.

[0171] The processor 1110 can include a video image compositor, which can be a processing block (e.g., implemented on a microprocessor), that implements video post-processing functions needed by the video playback application to produce the final image for the player window. The video image compositor can perform lens distortion correction on the wide-angle camera 1170, surround camera 1174, and / or on the cab-in monitor camera sensors. The cab-in monitor camera sensors are preferably monitored by a neural network running on another instance of the advanced SoC, configured to identify cab-in events and respond accordingly. The cab-in system can perform lip reading to activate mobile phone services and place a call, dictate an email, change the vehicle destination, activate or change the vehicle's infotainment system and settings, or provide voice-activated web surfing. Certain functions are only available to the driver when the vehicle is operating in autonomous mode, and are disabled otherwise.

[0172] The video image compositor can include enhanced temporal noise reduction for spatial and temporal noise reduction. For example, where motion is present in the video, the noise reduction appropriately weights the spatial information, reducing the weight of information provided by neighboring frames. Where the image or portions of the image do not include motion, the temporal noise reduction performed by the video image compositor can use information from previous images to reduce noise in the current image.

[0173] The video image compositor can also be configured to perform stereo correction on input stereo lens frames. The video image compositor can further be used for user interface composition when the operating system desktop is in use and the GPU 1108 does not need to continuously render new surfaces. Even when the GPU 1108 is powered on and active, doing 3D rendering, the video image compositor can be used to offload the GPU 1108 to improve performance and responsiveness.

[0174] The SoC 1104 can further include a Mobile Industry Processor Interface (MIPI) camera serial interface for receiving video and input from the cameras, a high-speed interface, and / or a video input block that can be used for camera and related pixel input functions. The SoC 1104 can further include an input / output controller that can be controlled by software and can be used to receive I / O signals that are not committed to a particular role.

[0175] The SoC 1104 can further include a wide range of peripheral device interfaces to enable communication with peripherals, audio codecs, power management, and / or other devices. The SoC 1104 can be used to process data from cameras (connected over Gigabit Multimedia Serial Link and Ethernet), sensors (e.g., LIDAR sensor 1164, RADAR sensor 1160, etc. that can be connected over Ethernet), data from the bus 1102 (e.g., speed of the vehicle 1100, steering wheel position, etc.), data from GNSS sensor 1158 (connected over Ethernet or CAN bus). The SoC 1104 can further include dedicated high performance mass storage controllers that can include their own DMA engines and that can be used to free up the CPU 1106 from routine data management tasks.

[0176] The SoC 1104 can be an end-to-end platform with a flexible architecture that spans automation levels 3-5, providing an integrated functional safety architecture for a platform that leverages and efficiently uses computer vision and ADAS technology to achieve diversity and redundancy, along with deep learning tools. The SoC 1104 can be faster, more reliable, and even more energy and space efficient than conventional systems. For example, the accelerators 1114, when combined with the CPU 1106, GPU 1108, and data storage 1116, can provide a fast and efficient platform for level 3-5 autonomous vehicles.

[0177] The technology thus provides capabilities and functionality that cannot be achieved by conventional systems. For example, computer vision algorithms can be executed on CPUs that can be configured using high-level programming languages such as the C programming language to perform a wide variety of processing algorithms across a wide variety of visual data. However, CPUs often cannot meet the performance requirements of many computer vision applications, such as those related to, for example, execution time and power consumption. In particular, many CPUs cannot execute complex object detection algorithms in real time, which is a requirement for on-board ADAS applications and for practical level 3-5 autonomous vehicles.

[0178] In contrast to conventional systems, by providing a CPU complex, a GPU complex, and a hardware acceleration cluster, the technology described herein allows multiple neural networks to be executed simultaneously and / or sequentially, and the results to be combined together to achieve level 3-5 autonomous driving functionality. For example, a CNN executed on a DLA or dGPU (e.g., GPU 1120) can include text and word recognition, allowing a supercomputer to read and understand traffic signs, including signs for which a neural network has not been specifically trained. The DLA can further include a neural network that is able to recognize, interpret, and provide a semantic understanding of the sign, and pass that semantic understanding to a path planning module running on the CPU complex.

[0179] As another example, multiple neural networks can be run simultaneously as required for level 3, 4, or 5 driving. For example, a warning sign consisting of the words "Caution: flashing lights indicate icy conditions" along with electric lights can be interpreted by several neural networks independently or collectively. The sign itself can be recognized by a first deployed neural network (e.g., a trained neural network) as a traffic sign, the text "flashing lights indicate icy conditions" can be interpreted by a second deployed neural network that informs the vehicle's path planning software (preferably executing on the CPU complex) that icy conditions exist when flashing lights are detected. The flashing lights can be recognized by operating a third deployed neural network over multiple frames that informs the vehicle's path planning software of the presence (or absence) of flashing lights. All three neural networks can be run simultaneously, for example, within the DLA and / or on the GPU 1108.

[0180] In some examples, a CNN for face recognition and owner recognition can use data from the camera sensors to recognize the presence of an authorized driver and / or owner of the vehicle 1100. A processing engine always on the sensor can be used to unlock the vehicle and turn on the lights when the owner approaches the driver's door, and in a safe mode, disable the vehicle when the owner leaves the vehicle. In this way, the SoC 1104 provides security against theft and / or carjacking.

[0181] In another example, a CNN for emergency vehicle detection and recognition can use data from the microphones 1196 to detect and recognize emergency vehicle sirens. In contrast to conventional systems that use a general classifier to detect sirens and manually extract features, the SoC 1104 uses a CNN to classify ambient and urban sounds as well as to classify visual data. In a preferred embodiment, a CNN running on the DLA is trained to recognize the relative closing speed of an emergency vehicle (e.g., by using the Doppler effect). The CNN can also be trained to recognize emergency vehicles specific to the local area in which the vehicle is operating as recognized by the GNSS sensor 1158. Thus, for example, when operating in Europe, the CNN will seek to detect European sirens, and when in the United States, the CNN will seek to recognize sirens that are only North American. Once an emergency vehicle is detected, a control program can be used to execute an emergency vehicle safety routine, slow the vehicle down, pull over to the side of the road, stop the vehicle, and / or idle the vehicle until the emergency vehicle passes, with the assistance of the ultrasonic sensors 1162.

[0182] The vehicle can include a CPU 1118 (e.g., a discrete CPU or dCPU) that can be coupled to the SoC 1104 via a high-speed interconnect (e.g., PCIe). The CPU 1118 can include, for example, an X86 processor. The CPU 1118 can be used to perform any of a wide variety of functions, including, for example, arbitrating potentially inconsistent results between ADAS sensors and the SoC 1104, and / or monitoring the status and health of the controller 1136 and / or infotainment SoC 1130.

[0183] The vehicle 1100 can include a GPU 1120 (e.g., a discrete GPU or dGPU) that can be coupled to the SoC 1104 via a high-speed interconnect (e.g., NVIDIA’s NVLINK). The GPU 1120 can provide additional artificial intelligence functionality, for example, by executing redundant and / or different neural networks, and can be used to train and / or update neural networks based at least in part on input (e.g., sensor data) from sensors of the vehicle 1100.

[0184] The vehicle 1100 can further include a network interface 1124 that can include one or more wireless antennas 1126 (e.g., one or more wireless antennas for different communication protocols, such as cellular antennas, Bluetooth antennas, etc.). The network interface 1124 can be used to enable wireless connections over the Internet with a cloud (e.g., with the server 1178 and / or other network devices), with other vehicles, and / or with computing devices (e.g., client devices of passengers). For communication with other vehicles, a direct link can be established between the two vehicles, and / or an indirect link can be established (e.g., across a network and through the Internet). The direct link can be provided using a car-to-car communication link. The car-to-car communication link can provide the vehicle 1100 with information about vehicles that are approaching the vehicle 1100 (e.g., vehicles in front of, to the side of, and / or behind the vehicle 1100). This functionality can be part of a cooperative adaptive cruise control functionality of the vehicle 1100.

[0185] The network interface 1124 can include a SoC that provides modulation and demodulation functionality and enables the controller 1136 to communicate over a wireless network. The network interface 1124 can include a radio frequency front end for up-conversion from baseband to radio frequency and down-conversion from radio frequency to baseband. The frequency conversion can be performed through well-known processes and / or can be performed using a super-heterodyne process. In some examples, the radio frequency front end functionality can be provided by a separate chip. The network interface can include wireless functionality for communication over LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-Wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0186] The vehicle 1100 can further include a data store 1128, which can include off-chip (e.g., off-SoC 1104) storage. The data store 1128 can include one or more storage elements, including RAM, SRAM, DRAM, VRAM, flash memory, hard disks, and / or other components and / or devices that can store data for at least one bit.

[0187] The vehicle 1100 can further include a GNSS sensor 1158. The GNSS sensor 1158 (e.g., GPS and / or assisted GPS sensor, differential GPS (DGPS) sensor, etc.) is used to assist in mapping, perception, occupancy grid generation, and / or path planning functions. Any number of GNSS sensors 1158 can be used, including, for example and without limitation, GPS using a USB connector with an Ethernet-to-serial (RS-232) bridge.

[0188] The vehicle 1100 can further include a RADAR sensor 1160. The RADAR sensor 1160 can be used by the vehicle 1100 for long-range vehicle detection, even in darkness and / or adverse weather conditions. The RADAR functional safety level can be ASIL B. The RADAR sensor 1160 can use the CAN and / or the bus 1102 (e.g., to transmit data generated by the RADAR sensor 1160) for control as well as access to object tracking data, in some examples, Ethernet for access to raw data. A wide variety of RADAR sensor types can be used. For example and without limitation, the RADAR sensor 1160 can be suitable for front, rear, and side RADAR use. In some examples, a pulsed Doppler RADAR sensor is used.

[0189] The RADAR sensor 1160 can include different configurations, such as long-range with narrow field of view, short-range with wide field of view, short-range side coverage, and so on. In some examples, long-range RADAR can be used for adaptive cruise control functionality. Long-range RADAR systems can provide a wide field of view (e.g., 250 m range) implemented through two or more independent scans. The RADAR sensor 1160 can help distinguish between static and moving objects, and can be used by the ADAS system for emergency brake assist and forward collision warning. The long-range RADAR sensor can include a single-station multi-mode RADAR with multiple (e.g., six or more) fixed RADAR antennas, as well as high-speed CAN and FlexRay interfaces. In examples with six antennas, the central four antennas can create focused beam patterns designed to record the surroundings of the vehicle 1100 at higher speed with minimal traffic interference from adjacent lanes. The other two antennas can extend the field of view, making it possible to quickly detect vehicles entering or leaving the lane of the vehicle 1100.

[0190] As one example, a mid-range RADAR system can include a range of up to 1160 m (front) or 80 m (rear) and a field of view of up to 42 degrees (front) or 1150 degrees (rear). A short-range RADAR system can include, but is not limited to, RADAR sensors designed to be mounted at both ends of the rear bumper. When mounted at both ends of the rear bumper, such a RADAR sensor system can create two beams that continuously monitor the rear and blind spots to the sides of the vehicle.

[0191] A short-range RADAR system can be used in an ADAS system for blind spot detection and / or lane change assist.

[0192] The vehicle 1100 can further include ultrasonic sensors 1162. The ultrasonic sensors 1162, which can be placed on the front, rear, and / or sides of the vehicle 1100, can be used for parking assist and / or to create and update an occupancy grid. A wide variety of ultrasonic sensors 1162 can be used, and different ultrasonic sensors 1162 can be used for different detection ranges (e.g., 2.5 m, 4 m). The ultrasonic sensors 1162 can operate at an ASIL B functional safety level.

[0193] The vehicle 1100 can include LIDAR sensors 1164. The LIDAR sensors 1164 can be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The LIDAR sensors 1164 can be at an ASIL B functional safety level. In some examples, the vehicle 1100 can include multiple LIDAR sensors 1164 (e.g., two, four, six, etc.) that can use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0194] In some examples, the LIDAR sensors 1164 can be capable of providing a list of objects and their distances for a 360-degree field of view. A commercially available LIDAR sensor 1164 can have, for example, an advertised range of approximately 1100 m, a precision of 2 cm - 3 cm, and support for a 1100 Mbps Ethernet connection. In some examples, one or more flush-mounted LIDAR sensors 1164 can be used. In such examples, the LIDAR sensors 1164 can be implemented as small devices that can be embedded into the front, rear, sides, and / or corners of the vehicle 1100. In such examples, the LIDAR sensors 1164 can provide a field of view of up to 120 degrees horizontal and 35 degrees vertical, with a range of 200 m, even for low reflectivity objects. Front-mounted LIDAR sensors 1164 can be configured for a horizontal field of view between 45 degrees and 135 degrees.

[0195] In some examples, LIDAR technology such as 3D Flash LIDAR can also be used. 3D Flash LIDAR uses a flash of laser light as a source of emission to illuminate the vehicle’s surroundings up to about 200 m. The flash LIDAR unit includes a receptor that records the laser pulse transmission time and reflected light on each pixel, which in turn corresponds to the range from the vehicle to the object. Flash LIDAR can allow for the generation of highly accurate and distortion-free images of the surroundings with each laser flash. In some examples, four flash LIDAR sensors can be deployed, one on each side of the vehicle 1100. Available 3D flash LIDAR systems include solid-state 3D staring array LIDAR cameras (e.g., non-scanning LIDAR devices) that have no moving parts other than a fan. The flash LIDAR device can use 5 nanosecond Class I (eye-safe) laser pulses per frame and can capture the reflected laser light in the form of 3D range point clouds and co-registered intensity data. By using flash LIDAR, and because flash LIDAR is a solid-state device with no moving parts, the LIDAR sensor 1164 can be less susceptible to motion blur, vibration, and / or jostling.

[0196] The vehicle can further include an IMU sensor 1166. In some examples, the IMU sensor 1166 can be located at the center of the rear axle of the vehicle 1100. The IMU sensor 1166 can include, for example and without limitation, an accelerometer, a magnetometer, a gyroscope, a magnetic compass, and / or other sensor types. In some examples, such as in six-axis applications, the IMU sensor 1166 can include an accelerometer and a gyroscope, while in nine-axis applications, the IMU sensor 1166 can include an accelerometer, a gyroscope, and a magnetometer.

[0197] In some embodiments, the IMU sensor 1166 can be implemented as a micro- high-performance GPS-aided inertial navigation system (GPS / INS) that combines micro-electromechanical systems (MEMS) inertial sensors, high-sensitivity GPS receivers, and advanced Kalman filtering algorithms to provide estimates of position, velocity, and attitude. As such, in some examples, the IMU sensor 1166 can enable the vehicle 1100 to estimate heading without input from a magnetic sensor by directly observing the change in velocity from GPS to the IMU sensor 1166 and correlating it. In some examples, the IMU sensor 1166 and the GNSS sensor 1158 can be combined into a single integrated unit.

[0198] The vehicle can include a microphone 1196 placed in and / or around the vehicle 1100. The microphone 1196 can be used for emergency vehicle detection and identification, among other things.

[0199] The vehicle can further include any number of camera types, including stereo cameras 1168, wide-view cameras 1170, infrared cameras 1172, surround-view cameras 1174, long and / or mid-range cameras 1198, and / or other camera types. These cameras can be used to capture image data around the entire periphery of the vehicle 1100. The types of cameras used depend on the embodiment and requirements of the vehicle 1100, and any combination of camera types can be used to provide the necessary coverage around the vehicle 1100. Further, the number of cameras can vary depending on the embodiment. For example, the vehicle can include six cameras, seven cameras, ten cameras, twelve cameras, and / or another number of cameras. As one example and not by way of limitation, the cameras can support Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Each of the cameras is described in more detail herein with respect to FIG. 11A and FIG. 11B are described in more detail.

[0200] The vehicle 1100 can further include vibration sensors 1142. The vibration sensors 1142 can measure vibrations of components of the vehicle, such as axles. For example, changes in vibration can indicate changes in the road surface. In another example, when two or more vibration sensors 1142 are used, differences between the vibrations can be used to determine the friction or slip of the road surface (e.g., when there is a difference in vibration between a power driven axle and a free spinning axle).

[0201] The vehicle 1100 can include an ADAS system 1138. In some examples, the ADAS system 1138 can include a SoC. The ADAS system 1138 can include adaptive / automatic / autonomous cruise control (ACC), cooperative adaptive cruise control (CACC), forward collision warning (FCW), automatic emergency braking (AEB), lane departure warning (LDW), lane keeping assist (LKA), blind spot warning (BSW), rear cross-traffic warning (RCTW), collision warning system (CWS), lane centering (LC), and / or other features and functionality.

[0202] The ACC system can use RADAR sensors 1160, LIDAR sensors 1164, and / or cameras. The ACC system can include longitudinal ACC and / or lateral ACC. Longitudinal ACC monitors and controls the distance to the vehicle immediately ahead of the vehicle 1100 and automatically adjusts the vehicle speed to maintain a safe distance from the vehicle ahead. Lateral ACC performs distance keeping and, if necessary, suggests a lane change for the vehicle 1100. Lateral ACC is related to other ADAS applications such as LCA and CWS.

[0203] CACC uses information from other vehicles, which can be received from other vehicles indirectly via a wireless link or through a network connection (e.g., through the Internet) via the network interface 1124 and / or the wireless antenna 1126. Direct links can be provided by a vehicle-to-vehicle (V2V) communication link, while indirect links can be an infrastructure-to-vehicle (I2V) communication link. Generally, the V2V communication concept provides information about the immediately preceding vehicles (e.g., vehicles immediately ahead of and in the same lane as the vehicle 1100), while the I2V communication concept provides information about traffic further ahead. A CACC system can include either or both of I2V and V2V information sources. Given information about vehicles ahead of the vehicle 1100, CACC can be more reliable, and it has the potential to improve traffic flow and reduce road congestion.

[0204] FCW systems are designed to alert the driver to a hazard so that the driver can take corrective action. FCW systems use a front-facing camera and / or RADAR sensor 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as displays, speakers, and / or vibrating components. FCW systems can provide warnings in the form of, for example, sound, visual warnings, vibrations, and / or quick brake pulses.

[0205] AEB systems detect an impending forward collision with another vehicle or other object and can automatically apply the brakes if the driver does not take corrective action within specified time or distance parameters. AEB systems can use a front-facing camera and / or RADAR sensor 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC. When an AEB system detects a hazard, it typically first alerts the driver to take corrective action to avoid a collision, and if the driver does not take corrective action, the AEB system can automatically apply the brakes in an effort to prevent or at least mitigate the effects of a predicted collision. AEB systems can include technologies such as dynamic brake support and / or crash imminent braking.

[0206] LDW systems provide visual, audible, and / or tactile warnings such as steering wheel or seat vibrations to alert the driver when the vehicle 1100 is crossing lane markers. The LDW system is not activated when the driver indicates an intentional lane departure by activating a turn signal. LDW systems can use a front-side facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as displays, speakers, and / or vibrating components.

[0207] An LKA system is a variation of the LDW system. If the vehicle 1100 begins to leave the lane, the LKA system provides a steering input or brake to correct the vehicle 1100.

[0208] A BSW system detects and warns the driver of vehicles in the car's blind spot. The BSW system can provide visual, audible, and / or tactile alerts to indicate that merging or changing lanes is unsafe. The system can provide additional warnings when the driver uses a turn signal. The BSW system can use rear-side facing cameras and / or RADAR sensors 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component.

[0209] A RCTW system can provide visual, audible, and / or tactile notifications when an object is detected outside the range of the rear-facing camera while the vehicle 1100 is backing up. Some RCTW systems include AEB to ensure that vehicle brakes are applied to avoid a collision. The RCTW system can use one or more rear-facing RADAR sensors 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component.

[0210] Conventional ADAS systems can be prone to false positive results, which can annoy and distract the driver, but typically are not catastrophic because the ADAS system alerts the driver and allows the driver to decide whether the safety condition is truly present and act accordingly. However, in an autonomous vehicle 1100, in the case of conflicting results, the vehicle 1100 itself must decide whether to heed the results from the primary computer or the secondary computer (e.g., the first controller 1136 or the second controller 1136). For example, in some embodiments, the ADAS system 1138 can be a secondary and / or auxiliary computer for providing perception information to a backup computer plausibility module. The backup computer plausibility monitor can run redundant diverse software on hardware components to detect faults in perception and dynamic driving tasks. The output from the ADAS system 1138 can be provided to a supervisory MCU. If the outputs from the primary and secondary computers conflict, the supervisory MCU must determine how to reconcile the conflict to ensure safe operation.

[0211] In some examples, the host computer can be configured to provide a confidence score to the supervisory MCU indicating the host computer's confidence in the selected result. If the confidence score exceeds a threshold, then the supervisory MCU can follow the host computer's direction, regardless of whether the secondary computer provides conflicting or inconsistent results. In the event that the confidence score does not satisfy the threshold and in the event that the host computer and the secondary computer indicate different results (e.g., a conflict), the supervisory MCU can arbitrate between the computers to determine the appropriate result.

[0212] The supervisory MCU can be configured to run a neural network that is trained and configured to determine conditions under which the secondary computer provides false alarms based at least in part on the output from the host computer and the secondary computer. Thus, the neural network in the supervisory MCU can learn when the output of the secondary computer can be trusted and when it cannot. For example, when the secondary computer is a RADAR-based FCW system, the neural network in the supervisory MCU can learn when the FCW system is identifying metal objects that are not in fact dangerous, such as drain grates or manhole covers that trigger false alarms. Similarly, when the secondary computer is a camera-based LDW system, the neural network in the supervisory MCU can learn to disregard the LDW when a cyclist or pedestrian is present and lane departure is in fact the safest strategy. In embodiments that include a neural network running on the supervisory MCU, the supervisory MCU can include at least one of a DLA or a GPU suitable for running a neural network with associated memory. In preferred embodiments, the supervisory MCU can include and / or be included as a component of the SoC 1104.

[0213] In other examples, the ADAS system 1138 can include a secondary computer that performs ADAS functions using traditional computer vision rules. In this way, the secondary computer can use classic computer vision rules (if-then), and the presence of a neural network in the supervisory MCU can improve reliability, safety, and performance. For example, the diverse implementation and intentional non-identity make the overall system more fault-tolerant, especially with respect to faults caused by software (or software-hardware interface) functions. For example, if there is a software bug or error in the software running on the host computer and the non-identical software code running on the secondary computer provides the same overall result, then the supervisory MCU can be more confident that the overall result is correct and that the bug in the software or hardware on the host computer did not cause a substantial error.

[0214] In some examples, the output of the ADAS system 1138 can be fed to a perception block of the host computer and / or a dynamic driving task block of the host computer. For example, if the ADAS system 1138 indicates a forward collision warning due to an object immediately ahead, the perception block can use this information in identifying the object. In other examples, the secondary computer can have its own neural network that is trained and thus reduces the risk of false positives as described herein.

[0215] The vehicle 1100 can further include an infotainment SoC 1130 (e.g., an in-vehicle infotainment system (IVI)). Although shown and described as a SoC, the infotainment system can not be a SoC and can include two or more discrete components. The infotainment SoC 1130 can include a combination of hardware and software that can be used to provide audio (e.g., music, a personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., TV, movies, streaming media, etc.), telephony (e.g., hands-free calling), network connectivity (e.g., LTE, Wi-Fi, etc.), and / or information services (e.g., a navigation system, a park assist, a radio data system, vehicle-related information such as a fuel level, a total distance covered, a brake fuel level, an oil level, a door open / close, air filter information, etc.) to the vehicle 1100. For example, the infotainment SoC 1130 can include a radio, a disc player, a navigation system, a video player, USB and Bluetooth connectivity, an in-car computer, in-car entertainment, Wi-Fi, steering wheel audio controls, hands-free voice controls, a heads-up display (HUD), an HMI display 1134, a telematics device, a control panel (e.g., for controlling and / or interacting with various components, features, and / or systems), and / or other components. The infotainment SoC 1130 can further be used to provide information (e.g., visual and / or audible) to a user of the vehicle, such as information from the ADAS system 1138, autonomous driving information such as planned vehicle maneuvers, trajectories, surrounding environment information (e.g., intersection information, vehicle information, road information, etc.), and / or other information.

[0216] The infotainment SoC 1130 can include GPU functionality. The infotainment SoC 1130 can communicate with other devices, systems, and / or components of the vehicle 1100 over a bus 1102 (e.g., a CAN bus, Ethernet, etc.). In some examples, the infotainment SoC 1130 can be coupled to a supervisory MCU such that, in the event of a failure of the host controller 1136 (e.g., a primary and / or backup computer of the vehicle 1100), the GPU of the infotainment system can perform some autonomous driving functions. In such examples, the infotainment SoC 1130 can place the vehicle 1100 in a driver safe park mode as described herein.

[0217] The vehicle 1100 can further include an instrument cluster 1132 (e.g., a digital dashboard, an electronic instrument cluster, a digital instrument panel, etc.). The instrument cluster 1132 can include a controller and / or supercomputer (e.g., a discrete controller or supercomputer). The instrument cluster 1132 can include a set of instruments, such as a speedometer, fuel level, oil pressure, tachometer, odometer, turn indicator, shift position indicator, seat belt warning light, parking brake warning light, engine malfunction light, supplemental restraint system (SRS) system information, lighting controls, safety system controls, navigation information, etc. In some examples, information can be displayed and / or shared between the infotainment SoC 1130 and the instrument cluster 1132. In other words, the instrument cluster 1132 can be included as part of the infotainment SoC 1130, or vice versa.

[0218] FIG. 11D FIG. 11 illustrates a system diagram of communication between a cloud-based server and an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. FIG. 11A FIG. 11 illustrates a system diagram of communication between a cloud-based server and an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. The system 1176 can include servers 1178, a network 1190, and vehicles including the vehicle 1100. The servers 1178 can include a plurality of GPUs 1184(A)- 1184(H) (collectively referred to herein as GPUs 1184), PCIe switches 1182(A)- 1182(H) (collectively referred to herein as PCIe switches 1182), and / or CPUs 1180(A)- 1180(B) (collectively referred to herein as CPUs 1180). The GPUs 1184, CPUs 1180, and PCIe switches can be interconnected with high-speed interconnects such as, for example and without limitation, NVLink interfaces 1188 developed by NVIDIA and / or PCIe connections 1186. In some examples, the GPUs 1184 are connected via NVLink and / or NVSwitch SoC connections, and the GPUs 1184 and PCIe switches 1182 are connected via PCIe interconnects. Although eight GPUs 1184, two CPUs 1180, and two PCIe switches are shown, this is not intended to be limiting. Depending on the embodiment, each of the servers 1178 can include any number of GPUs 1184, CPUs 1180, and / or PCIe switches. For example, each of the servers 1178 can include eight, sixteen, thirty-two, and / or more GPUs 1184.

[0219] The server 1178 can receive image data from vehicles over the network 1190 and representing images showing unexpected or changing road conditions such as a road work that has recently started. The server 1178 can transmit neural networks 1192, updated neural networks 1192, and / or map information 1194, including information about traffic and road conditions, to vehicles over the network 1190. Updates to the map information 1194 can include updates to the HD map 1122, e.g., information about construction sites, potholes, curves, flooding, or other obstacles. In some examples, the neural networks 1192, updated neural networks 1192, and / or map information 1194 can have been produced from experience based at least in part on training performed at a data center (e.g., using the server 1178 and / or other servers) and / or from data received from any number of vehicles in the environment.

[0220] The server 1178 can be used to train machine learning models (e.g., neural networks) based at least in part on training data. The training data can be generated by vehicles and / or can be generated in simulations (e.g., using game engines). In some examples, the training data is labeled (e.g., in cases where the neural network benefits from supervised learning) and / or undergoes other pre-processing, while in other examples, the training data is not labeled and / or pre-processed (e.g., in cases where the neural network does not require supervised learning). The training can be performed according to any class or more classes of machine learning techniques, including but not limited to the following classes: supervised training, semi-supervised training, unsupervised training, self-learning, reinforcement learning, federated learning, transfer learning, feature learning (including principal component and cluster analysis), multilinear subspace learning, manifold learning, representation learning (including spare dictionary learning), rule-based machine learning, anomaly detection, and any variants or combinations thereof. Once the machine learning models are trained, the machine learning models can be used by vehicles (e.g., transmitted to vehicles over the network 1190) and / or the machine learning models can be used by the server 1178 to remotely monitor vehicles.

[0221] In some examples, the server 1178 can receive data from vehicles and apply the data to the latest real-time neural networks for real-time intelligent inference. The server 1178 can include deep learning supercomputers and / or specialized AI computers powered by GPUs 1184, such as the DGX and DGX Station machines developed by NVIDIA. However, in some examples, the server 1178 can include deep learning infrastructure of a data center that is powered by CPUs only.

[0222] The deep learning infrastructure of the server 1178 can be capable of fast real-time inference, and can use this capability to assess and validate the health of the processors, software, and / or associated hardware in the vehicle 1100. For example, the deep learning infrastructure can receive periodic updates from the vehicle 1100, such as a sequence of images and / or objects located in the sequence of images that the vehicle 1100 has located (e.g., via computer vision and / or other machine learning object classification techniques). The deep learning infrastructure can run its own neural network to identify the objects and compare them to the objects identified by the vehicle 1100, and if the results do not match and the infrastructure concludes that the AI in the vehicle 1100 is malfunctioning, the server 1178 can transmit a signal to the vehicle 1100 instructing the fail-safe computer of the vehicle 1100 to take control, notify the passengers, and complete a safe parking operation.

[0223] For inference, the server 1178 can include GPUs 1184 and one or more programmable inference accelerators (such as NVIDIA’s TensorRT). The combination of GPU-powered servers and inference-accelerated can make real-time response possible. In other examples, such as where performance is less important, CPU, FPGA, and other processor-powered servers can be used for inference.

[0224] The present disclosure can be described in the general context of machine-usable instructions or computer code, including computer-executable instructions such as program modules, being executed by a computer or other machine, such as a personal digital assistant or other handheld device. Generally, program modules including routines, programs, objects, components, data structures, and the like, refer to code that performs particular tasks or implements particular abstract data types. The present disclosure can be practiced in a variety of system configurations, including hand-held devices, consumer electronics, general- purpose computers, more specialty computing devices, and the like. The present disclosure can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network.

[0225] As used herein, the term “and / or” placed between a first element and a second element can have the meaning interpreted as the following example: the first element or the second element or the combination of the first element and the second element. For example, “element A and / or element B” can include only element A, only element B, or both element A and element B. Additionally, the phrase “at least one of element A or element B” can include at least one of element A, at least one of element B, or at least one of element A and at least one of element B. Further, the phrase “at least one of element A and element B” can include at least one of element A, at least one of element B, or at least one of element A and at least one of element B.

[0226] The subject matter of the present disclosure is described with specificity herein to meet statutory requirements. However, the description itself is not intended to limit the scope of this disclosure. Rather, the inventors have contemplated that the claimed subject matter might also be embodied in other ways, to include different steps or combinations of steps similar to the ones described in this document, in conjunction with other present or future technologies. Moreover, although the terms "step" and / or "block" might be used herein to connote different elements of methods employed, the terms should not be interpreted as implying any particular order among or between various steps herein disclosed unless and except when the order of individual steps is explicitly described.

Claims

1. A method for training a neural network, comprising: receiving images of one or more objects with a plurality of backgrounds; generating a set of inference scores corresponding to a first set of inference data, the scores corresponding to one or more predictions of a prediction task performed using the images on the one or more objects; selecting a background based at least on one or more inference scores of the set of inference scores; generating an image based at least on integrating an object with the background based at least on the selection of the background; applying the image during training of at least one neural network to perform the prediction task and generate a second set of inference data; and updating the at least one neural network using the second set of inference data.

2. The method of claim 1, wherein the one or more inference scores comprise a plurality of inference scores for a set of the images comprising the background, and the selection is based at least on an analysis of the plurality of inference scores.

3. The method of claim 1, wherein the inference scores are generated using the at least one neural network in a first epoch of training the at least one neural network and the use of the image is in a second epoch of the training.

4. The method of claim 1, wherein generating the image comprises generating a mask based at least on identifying a region of the object in the image and applying the mask to the image.

5. The method of claim 1, wherein the selection of the background is based at least on determining that at least one of the one or more inference scores is below a threshold.

6. The method of claim 1, wherein the one or more inference scores correspond to a first cropped region of the background and the object is integrated with a second cropped region of the background, the second cropped region of the background being different than the first cropped region.

7. The method of claim 1, wherein the background is a background type and the method further comprises synthetically generating the background based at least on the background type.

8. A system for training a neural network, comprising: one or more processors; and one or more memory devices storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising: receiving images of one or more objects with a plurality of backgrounds; generating a set of inference scores corresponding to a first set of inference data, the scores corresponding to one or more predictions of a prediction task performed using the images on the one or more objects; selecting a background based at least on one or more inference scores of the set of inference scores; generating an image based at least on integrating an object with the background based at least on the selection of the background; applying the image during training of at least one neural network to perform the prediction task and generate a second set of inference data; and updating the at least one neural network using the second set of inference data. ​ ​ ​ 9. The system of claim 8, wherein the one or more inference scores comprise a plurality of inference scores for a set of the images that include the background, and the selection is based at least on an analysis of the plurality of inference scores.

10. The system of claim 8, wherein the inference score is generated using the at least one neural network in a first epoch of training the at least one neural network and the use of the image is in a second epoch of the training.

11. The system of claim 8, wherein generating the image comprises generating a mask based at least on identifying a region of the object in the image and applying the mask to the image.

12. The system of claim 8, wherein the selection of the background is based at least on determining that at least one of the one or more inference scores is below a threshold.

13. The system of claim 8, wherein the one or more inference scores correspond to a first cropping region of the background and the object is integrated with a second cropping region of the background that is different than the first cropping region.

14. The system of claim 8, wherein the background is a background type and the background is synthetically generated based at least on the background type.

15. The system of claim 8, wherein the operations are performed by at least one of: a control system of an autonomous or semi-autonomous machine; a perception system of an autonomous or semi-autonomous machine; a system for performing simulation operations; a system that performs deep learning operations; a system implemented using edge devices; a system implemented using robots; a system that consolidates one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources.

Images