Training of neural network for more reliably detecting objects even if they are of unknown type

Abstract

To provide a method for training a neural network.SOLUTION: A method in a neural network 1 that extracts features from images by using a feature extractor network and determines, from these features, classification scores with respect to one or more classes out of a given set of classes by using a classifier head, includes: a step 110 of providing training images 2a and respective ground truth classification scores 5a; a step 120 of processing these training images or regions thereof into classification scores by using the neural network; and a step 130 of calculating a value 7a of a loss function 7. The loss function is dependent at least on a deviation between the classification scores and the ground truth classification scores and object likeliness contribution that is dependent on the presence or absence of an object but not dependent on class information. The method further includes a step 140 of optimizing parameters 1a that characterize the behavior of the neural network toward the goal of improving the value of the loss function.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to image classification, and in particular to detecting and / or locating objects of known and unknown types within images, which detection is of great importance for safety-related applications such as the autonomous operation of vehicles or robots. [Background technology]

[0002] background In order for a vehicle or robot to operate autonomously within a company's premises or even in public road traffic, it is necessary to constantly monitor the surroundings of the vehicle and / or robot. A crucial part of such monitoring is acquiring and analyzing images of such surroundings. It is particularly important to detect objects that the vehicle and / or robot may collide with.

[0003] An object detector locates objects of interest and associates a classification score for one or more classes of a given classification with each detected object instance. That is, the object detector can attribute some images to a particular type of object. However, object detectors have difficulty reliably detecting out-of-distribution (OOD) objects, i.e., objects not seen during training or that differ significantly from known categories. An example of such an object is cargo such as skis or furniture dropped by another vehicle on the road. Summary of the Invention [Problem to be solved by the invention]

[0004] Disclosure of the Invention The present invention provides a method for training a neural network. The neural network is configured to extract features from an image using a feature extractor network. A classifier head of the neural network then uses these features to determine a classification score for one or more classes of a given set of classes. For example, the feature extractor may include one or more convolutional layers, and the classifier head may include one or more fully connected layers. On top of the classifier head, the neural network may include other heads that utilize the features, such as one or more regressor heads that determine the value of any desired quantity. [Means for solving the problem]

[0005] During the method, training images and their respective ground truth classification scores are provided. These training images or regions thereof are processed into classification scores using a neural network. That is, an object detector can first detect regions of interest in the images that may indicate the presence of an object instance, and then the neural network can map each region of interest to a classification score. To estimate how good the output of the neural network is, a loss function value is calculated. This loss function takes into account at least two contributions: The deviation between the classification score and the ground truth classification score; an objectness contribution that depends on the presence or absence of an object but does not depend on class information; In particular, this object-likeness contribution can relate to a particular region of interest within the image, rather than to the entire image.

[0006] The object-likeness contribution may, for example, be determined from the classification scores, although this is not required: rather, the object-likeness contribution may be determined from other outputs of the neural network, such as from the output of a further head dedicated to object-likeness, or even directly from the features.

[0007] The parameters that characterize the behavior of the neural network are optimized with the goal of improving the value of the loss function.

[0008] The inventors have discovered that including the object-likeness contribution in the loss improves the reliability of object detection. In particular, reliability can include correctly identifying pixels or other portions of an image that belong to an object when the image shows the presence of the actual object at a particular location. Reliability can also include not detecting a "ghost" object when the image does not show the actual object. The relative importance of these goals depends on the application at hand. For example, in autonomous driving applications, it is crucial that the presence of an object, when present, is detected so that a collision between the autonomous vehicle or robot and the object can be avoided. However, in public road traffic, it is also crucial that there are no false detections of objects that are not actually present. False detections can trigger emergency braking or evasive maneuvers that are not actually justified. Therefore, such maneuvers occur as a complete surprise to other traffic participants, which in turn can cause a subsequent collision.

[0009] Also, improved detection of out-of-distribution, unseen objects is particularly important for quickly reacting to completely unexpected situations in road traffic. For example, it is extremely rare to encounter a cow, a bench, or skis on a highway. However, if a bench or skis is really present because someone dropped it during transport, or if a cow is present because it has escaped from a farm, it is important to initiate emergency braking and / or avoidance.

[0010] At the same time, the independence of the object-likeness contribution from class information differs from using only class-independent cues. Class-independent object-likeness contributions can take advantage of features whose specific combinations indicate a class. For example, when characterizing a person, a specific combination of feature values ​​for “height,” “build,” and “facial expression” may indicate the class “investment banker,” while another combination of these feature values ​​may indicate the class “aggressive criminal.” Object-likeness analysis can fully use the “height,” “build,” and “facial expression” features to determine where a person is present in an image independently of the person’s potential class. In contrast, using only class-independent cues may mean that the “height,” “build,” and “facial expression” features are excluded from the analysis of whether a person is present. This can be undesirable, because, for example, a person walking on a road between two parked cars may be partially occluded by the parked cars in the image, with only their face visible. In this case, the feature of "facial expression" would be the only cue that would allow a person to be detected in time before they jump out into the road to the right of the automatically controlled vehicle.

[0011] In particular, the independence of the object-likeness contribution from class information ensures that the neural network truly learns the concept of object-likeness as such in a general sense, rather than simply augmenting the set of classes it knows with a few classes. Simply exposing the neural network to outliers that do not belong to the original set of classes still cannot guarantee this; such exposure to outliers may only trigger learning. Also, the inference of the trained network is not slowed down, as may be the case, for example, when simply adding anomaly detection to an existing neural network. Thus, real-time performance, which is important for autonomous driving and other time-critical applications, is not hindered.

[0012] In terms of general object-likeness, objects typically contain multiple well-connected surfaces and have a specific geometric structure. These cues are common across different class categories. Learning these cues to detect the presence of an object allows generalization from known classes to unknown classes during inference. If the cues were class-specific, this would not be possible because the object-likeness would be biased toward only known classes.

[0013] Although learning object-likeness during training requires some effort, the ability to detect object-likeness does not necessarily impose a significant computational burden during inference. This is particularly important for automotive and other mobile applications where hardware resources are limited while at the same time a fast decision regarding the presence or absence of an object is required. That is, during inference, only a limited amount of additional resources can be allocated to the additional capability of unknown object detection.

[0014] Furthermore, learning object-likeness does not degrade the performance of the neural network for known classes within a given set of classes. For example, if a contribution related to the classification score and an object-likeness contribution are added to the loss function, even if the object-likeness contribution is too low for a particular training image, it cannot offset the large value of the contribution related to the classification score. That is, the neural network cannot avoid the burden of becoming better at determining classification scores by becoming better at determining object-likeness.

[0015] In a particularly advantageous embodiment, the object-likeness contribution depends on the output of a further object-likeness head of the neural network, which predicts at least one occupancy in a class-independent manner. This occupancy is a measure of whether a feature suggests the presence of an object. This occupancy is then translated into whether a particular region in the image suggests the presence of an actual object. In one example, the object-likeness head may include several convolutional layers with nonlinear activation functions between them. Such an object-likeness head may predict a single logit value. For example, a sigmoid mapping may then map this single logit value to a value between 0 and 1. In particular, the presence of a dedicated object-likeness head provides an additional possibility to ensure that training on object-likeness does not degrade performance for known classes. For example, training on object-likeness may be limited to optimizing parameters characterizing the behavior of the object-likeness head, while parameters characterizing the behavior of the classifier head and the feature extractor remain frozen. The presence of a separate object-likeness head also ensures that the classifier head outputs classification scores for a given known class without additional delay. That is, the object-likeness is determined purely based on the classification score.

[0016] Alternatively, or in combination with determining occupancy using a dedicated object-likeness head, occupancy can be determined from the output of a regressor in a neural network. For example, the YOLOX architecture: A feature pyramid network (FPN) as a feature extractor that outputs a set of feature maps; A regressor head that maps features from the feature map to bounding box coordinates of detected object instances; A classifier head that maps features from the feature map to class labels for detected object instances; an object-likeness head that uses intermediate work products from the regressor head to predict class-independent object-likeliness, and optionally can also use features from the classifier head; That is, for example, features from the bounding box regression can be fused with features from the classifier head.

[0017] Therefore, in a further particularly advantageous embodiment, the neural network is further configured to predict bounding boxes for objects, such bounding boxes corresponding to object instances providing another notion of object-likeness that is considered relevant to the output of the object-likeness head.

[0018] Therefore, in a further particularly advantageous embodiment, the object-likeness contribution depends on how well the occupancy matches one or more intersections between the predicted bounding box and the ground truth bounding box. That is, the localization of the predicted bounding box can be utilized to obtain a ground truth loss for predicting the occupancy. If the overlap between the predicted bounding box and the annotation of the ground truth bounding box is large, this indicates that the occupancy o should be high; otherwise, the occupancy o should be low. For example, the occupancy o can be calculated using the following formula:

number

[0019] Notably, one advantage of this is that detection of unknown objects does not depend on portions of the image suggesting the presence of an object sufficient for positive identification by matching with a particular class of ground truth. That is, even if an object is partially occluded or otherwise difficult to recognize, it is still possible to detect that at least some object is present. For example, if a vehicle is partially occluded, it may be difficult to distinguish the specific type of vehicle (e.g., car, van, or make and model). Also, if a person is partially visible among parked cars, it does not matter whether the specific type of person can be identified. What matters is detecting the presence of a person, or more abstractly, at least one object that should not be run over.

[0020] One way to measure the agreement between this representation on the one hand and the occupancy o on the other hand is the cross-entropy, e.g., the binary cross-entropy BCE. Hence, the object-likeness contribution to the loss, L occ teeth,

number

[0021] In information theory, cross-entropy is a measure of the quality of a model of a (probability) distribution. Hence, optimizing model parameters with the goal of minimizing the cross-entropy works towards maximizing the log-likelihood of the model given the distribution.

[0022] One particular advantage of having an occupancy score o is that this occupancy score o can be determined for all images. That is, all available training images can be used to train the determination of the occupancy score o, not just training images outside of known classes in a given set of classes. In contrast, training to determine classification scores for newly introduced OOD classes is very likely to overfit to the available OOD examples, since the number of available OOD examples is much smaller than the number of training examples in the distribution.

[0023] L occ In the formula, the intersection union is calculated as

number

[0024] That is, the predicted bounding box b p and all ground truth bounding boxes b gi Rather than calculating the rather complicated intersection between the union of b p and the individual ground truth bounding boxes b gi For most of these intersections, it is immediately possible to determine that they are empty without going too deep into the calculations, resulting in substantial savings in computation time.

[0025] Therefore, in a further particularly advantageous embodiment, the intersection between the predicted bounding box and the union of multiple ground truth bounding boxes is approximated by the sum of the intersections between the predicted bounding box and each ground truth bounding box.

[0026] In a further particularly advantageous embodiment, the given set of classes is extended with an additional class for objects that do not belong to any class within the given set of classes. In this way, the neural network classifier head has the opportunity to express the finding that a detected object is an unseen object. That is, the output of the classifier head can differentiate between "not an object" on the one hand and "object, but unseen object" on the other hand. Without the additional OOD classes, the classifier head might express both "not an object" and "object, but unseen object" with low scores for all given classes, or the classifier head might even be led to output high classification scores for any of the given classes that are all incorrect. Furthermore, the additional classification scores for the additional classes of unseen objects are obtained during estimation with only a small, if any, additional computational load.

[0027] In a further particularly advantageous embodiment, the set of training images is augmented with training images that do not belong to any class within the given set of classes. In this way, the neural network has an improved chance of detecting unseen objects. The additional training images for this augmentation can come from any suitable source. For example, to expose the neural network to unseen objects, multiple images from different data sets can be assembled into a single image using any known augmentation technique, such as Mosaic or Mixup. Exposure to a wide variety of objects enhances the acquisition of a more general understanding of object-likeness and may be performed in any suitable manner. For example, training images from other data sets can be used, and these training images can be further modified by any suitable data augmentation technique. In one example, a data set with training images of traffic situations for autonomous driving can be augmented with additional training images from the general-purpose MS COCO (Microsoft Common Objects in Context) large-scale object detection, segmentation, and captioning data set and / or the LVIS data set for large vocabulary instance segmentation. In particular, this makes the object-likeness scores more likely to respond to objects from both known and unknown classes, while remaining silent about "thing" classes such as roads and skies.

[0028] In a further particularly advantageous embodiment, images acquired by the at least one sensor are processed by a trained machine learning model into classification scores, and optionally also into occupancy and / or object-likeness scores, where the improved training has the effect of making object detection more reliable, whether the objects are of classes within an originally given set of classes or of classes outside these classes.

[0029] In a further particularly advantageous embodiment, in response to the classification score and / or occupancy indicating the presence of an object, the presence of the object is verified using depth information. For this purpose, depth information for an image region associated with the object is obtained. Then, it is determined whether the depth information indicates a depth change that would be expected if the object were present. If the determination is negative, i.e., if the expected depth change is not present, the object detection is determined to be a false positive. In particular, if an image displays features that somehow have the appearance of an object but do not belong to the actual object, these features will not be falsely detected as an object. An example of such a feature is a shadow. A shadow is generated by the presence of an actual object, but appears in a different location where no object is present. Examples of depth changes that indicate the presence of an object include the absence of local depth changes within a bounding box associated with the object. In contrast, for example, the flat surface of a road only shows continuous local depth changes.

[0030] In a further particularly advantageous embodiment, the classification score and the occupancy rate are evaluated together to verify the presence of an object. For this purpose, if the classification score and / or the occupancy rate suggest the presence of an object, the product of the highest classification score and the occupancy rate associated with the detected object is calculated. If this product is below a predetermined threshold, the detection of the object is determined to be a false positive. As mentioned above, this works best when there is an additional class for objects that do not belong to any class within a given set of classes. The presence of the object can then be confirmed by two independent heads, namely, the classifier head and the object-likeness head, after which it is concluded that the object actually exists.

[0031] In a further particularly advantageous embodiment, an operating signal is calculated based at least in part on the classification score and / or occupancy rate output by the trained machine learning model and / or based at least in part on the detection of the object. A vehicle, a robot, a driver assistance system, a quality inspection system, a monitoring system, and / or a medical imaging system are then operated by the operating signal. In this way, the probability that the reaction of the respective operated system to the operating signal will be appropriate for the situation characterized by the acquired image is improved. In particular, the absence of a reaction that should be performed in response to the actual presence of an object is reduced, and the implementation of a reaction in response to the detection of an object that does not correspond to an actually present object is reduced. For example, in an automated driving system, an emergency braking maneuver or an evasive maneuver is more reliably initiated when an object is actually present on the vehicle's path, but an emergency braking maneuver or an evasive maneuver is not "suddenly" performed for no apparent reason when an object is not actually present on the vehicle's path.

[0032] The method may be wholly or partly computer-implemented and may be embodied in software. Accordingly, the present invention also relates to a computer program comprising machine-readable instructions that, when executed by one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the above-described method. In this specification, control units for vehicles and control units for other embedded systems capable of executing machine-readable instructions should also be considered as computers. Computing instances include virtual machines, containers, or other execution environments that enable the execution of machine-readable instructions in the cloud.

[0033] The non-transitory storage medium and / or downloadable product may include a computer program. A downloadable product is an electronic product that can be sold online for immediate realization and transmitted over a network. One or more computers and / or computing instances may implement such computer program and / or such non-transitory storage medium and / or downloadable product.

[0034] In the following, the invention will be described with the aid of the drawings, without any intention of limiting the scope of the invention. [Brief explanation of the drawings]

[0035] [Figure 1] FIG. 1 illustrates an exemplary embodiment of a method 100 for training a neural network 1. [Figure 2] FIG. 1 is a visualization of an exemplary processing pipeline for training and an exemplary processing pipeline for inference according to method 100. [Figure 3] FIG. 10 is a diagram illustrating an example of an image that causes erroneous object detection. DETAILED DESCRIPTION OF THE INVENTION

[0036] FIG. 1 illustrates a method 100 for training a neural network 1, the neural network 1 comprising: Extract features 3 from image 2 using a feature extractor network 4; From these features 3, a classifier head 6 is used to determine a classification score 5 for one or more classes of a given set of classes. 1 is a schematic flow chart of an exemplary embodiment of a method 100 configured to:

[0037] According to block 105 , the neural network 1 may be further configured to predict a bounding box 10 for an object detected in the image 2 .

[0038] In step 110, training images 2a and their respective ground truth classification scores 5a are provided.

[0039] According to block 111, the set of training images 2a can be extended by training images 2a* that do not belong to any class within the given set of classes.

[0040] In step 120, training images 2a are processed using the neural network 1 into classification scores 5.

[0041] In step 130, a value 7a of a given loss function 7 is calculated. The loss function 7 has at least the deviation between the classification score 5 and the ground truth classification score 5a; The object-likeness contribution depends on the existence or non-existence of an object but does not depend on class information. Depends on.

[0042] According to block 131, the object-likeness contribution may depend on the output of a further object-likeness head 8 of the neural network 1, which predicts an occupancy 9 in a class-independent manner. This occupancy 9 is a measure of whether the feature 3 is suggestive of the presence of an object.

[0043] According to block 132, the object-likeness contribution may depend on how well the occupancy 9 matches one or more intersections between the predicted bounding box 10 and the ground truth bounding box 10a.

[0044] According to block 132a, the intersection between the predicted bounding box 10 and the union of multiple ground truth bounding boxes 10a can be approximated by the sum of the intersections between the predicted bounding box 10 and each of the ground truth bounding boxes 10a.

[0045] According to block 132b, the agreement between the occupancy 9 and the intersection can be measured by cross-entropy.

[0046] According to block 133, the given set of classes can be extended with further classes for objects that do not belong to any class in the given set of classes.

[0047] In step 140, parameters 1a characterizing the behavior of neural network 1 are optimized with the goal of improving the value 7a of loss function 7. The final optimized state of the parameters is labeled with reference sign 1a* and characterizes the trained state 1* of neural network 1.

[0048] In the example shown in FIG. 1, in step 150, images 2 acquired by at least one sensor 11 are processed by a trained machine learning model 1* into a classification score 5, and optionally also into an occupancy 9 and / or an object-likeness score.

[0049] In step 160, it is checked whether the classification score5 and / or occupancy9 and / or object-likeness score suggest the presence of an object. If this is the case (truth value 1), false positive detections can be removed by one or both of the following approaches, which may also be performed iteratively for multiple instances of the detected object.

[0050] According to the first approach, depth information 12 for an image region associated with an object (here labeled O) is obtained in step 170. It is then determined in step 180 whether this depth information 12 indicates a depth change that may be expected if the object is present. If this is not the case (truth value 0), then in step 190 the detection of object O is determined to be a false positive.

[0051] According to the second approach, in step 200, the product 13 of the best classification score 5 and the occupancy 9 associated with this detected object O is calculated. In step 210, it is checked whether this product 13 is above a predetermined threshold 14. If this is not the case (truth value 0), in step 220, the detection of object O is determined to be a false positive.

[0052] In step 230, an operation signal 230a is calculated based at least in part on the classification scores 5 and / or occupancy rates 9 output by the trained machine learning model 1* and / or based at least in part on the detection of the object O. In step 240, the vehicle 50, the driver assistance system 51, the robot 60, the quality inspection system 70, the surveillance system 80, and / or the medical imaging system 90 are operated by the operation signal 230a.

[0053] FIG. 2 shows an exemplary processing pipeline for training and an exemplary processing pipeline for inference according to the method 100 described above.

[0054] Figure 2(a) shows the pipeline for training. According to block 111, training images 2a from the domain of the current application (here, autonomous driving) are combined with further training images 2a* from outside this domain. The combined set of training images is fed to a feature extractor 4 of the neural network 1. This feature extractor 4 produces extracted features 3.

[0055] The extracted features 3 are fed to a classifier head 6 which generates classification scores 5 for in-distribution class IDs from a given set of classes, as well as for a new out-of-distribution class OOD for objects of unknown classes.

[0056] The extracted features 3 are also fed to an object-likeness head 8, which predicts at least one occupancy 9 in a class-independent manner, which is a measure of whether the features 3 suggest the presence of an object. In the example shown in Figure 2, the object-likeness head 8 is further configured to determine an "obj" score, which is a further notion of object-likeness. The object-likeness head 8 is also used to predict a bounding box 10 for the object.

[0057] In the example shown in FIG. 2, the training of neural network 1 targets the following objectives: ·Loss component L box that the predicted bounding box 10 corresponds to the ground truth bounding box 10a, as measured by ·Loss component L occ The occupancy 9 corresponds to the bounding box 10, 10a, as measured by ·Loss component L obj that the object score "obj" matches the available ground truth, as measured by any suitable technique by ·Classification loss component L cls that the classification score 5 corresponds to the ground truth classification score 5a, as measured by The purpose is to:

[0058] Figure 2(b) starts with the assumption that a neural network 1 has been trained according to the pipeline shown in Figure 2(a). Figure 2(b) shows an exemplary inference pipeline. The trained neural network 1* is fed an image 2. The neural network 1* then outputs both an occupancy map 9 of the image 2 and a classification score 5 visualized by bounding boxes 10 for in-distribution objects of known classes (ID) and for out-of-distribution objects of unknown classes (OOD).

[0059] In the example shown in FIG. 2(b), occupancy map 9 has a high occupancy score for one exemplary object O whose presence is not suggested by classification score 5. Classification score 5 also suggests the presence of one other object O' that is not evident from occupancy map 9. In step 200, the product 13 of classification score 5 and occupancy map 9 is calculated, and object O is determined to be present if this product is above threshold 14. In the example shown in FIG. 2, the total set of detected objects O is the union of both the objects detected using classification score 5 and the objects detected using occupancy map 9.

[0060] 2(c) shows how these objects O are further filtered by the depth information 12 according to steps 180 and 190 of method 100. In the example shown in FIG. 2, for object O, matching depth information 12 exists, so this object O is retained. In contrast, for object O', matching depth information 12 is not available. Therefore, this object O' is discarded as a false positive.

[0061] Figure 3 shows some examples of road scene images 2 that cause false object detections. The bounding boxes for these false detections are shown with dashed lines.

[0062] In Figure 3(a), taken from the Fishyscapes dataset, a manhole cover flush with the road surface and graffiti painted on the road surface cause false positives.

[0063] In Figure 3(b), taken from the Cityscapes dataset, subtle texture changes in the road resulting from road surface repairs cause false positives.

[0064] In Figure 3(c), obtained from the BDD100K dataset, road markings cause false positives.

Claims

1. A method (100) for training a neural network (1), comprising: The neural network (1) is configured to extract features (3) from an image (2) using a feature extractor network (4) and to determine classification scores (5) from the features (3) for one or more classes of a given set of classes using a classifier head (6); The method (100) comprises: Providing (110) training images (2a) and their respective ground truth classification scores (5a); - processing (120) the training images (2a) or regions thereof using the neural network (1) into classification scores (5); A step (130) of calculating a value (7a) of a loss function (7), said loss function (7) comprising at least the deviation between the classification score (5) and the ground truth classification score (5a); ○ An object-likeness contribution that depends on the presence or absence of an object but does not depend on class information; Step (130), which depends on a step (140) of optimizing the parameters (1 a) characterizing the behavior of said neural network (1) with the goal of improving said value (7 a) of said loss function (7); A method (100) comprising:

2. said object-likeness contribution depends (131) on the output of a further object-likeness head (8) of said neural network (1), which predicts at least one occupancy (9) in a class-independent manner; The occupancy (9) is a measure of whether the feature (3) indicates the presence of an object. The method (100) of claim 1.

3. The neural network (1) is further configured to predict (105) a bounding box (10) for an object.

3. The method (100) of claim 1 or 2.

4. The object-likeness contribution depends on how well the occupancy (9) matches (132) one or more intersections between the predicted bounding box (10) and a ground truth bounding box (10a). The method (100) of claims 2 and 3.

5. The intersection between the predicted bounding box (10) and a union of a plurality of ground truth bounding boxes (10a) is approximated (132a) by the sum of the intersections between the predicted bounding box (10) and each ground truth bounding box (10a). The method (100) of claim 4.

6. The agreement between the occupancy (9) and the intersection is measured by cross-entropy (132b). The method (100) according to claim 4 or 5.

7. The given set of classes is extended (133) with a further class for objects that do not belong to any class in the given set of classes. The method (100) of any one of claims 1 to 6.

8. The set of training images (2a) is augmented (111) with training images (2a*) that do not belong to any class within the given set of classes. The method (100) of claim 7.

9. The method (100) comprises: processing (150) images (2) acquired by at least one sensor (11) by a trained machine learning model (1*) into classification scores (5), and optionally also into occupancy (9) and / or object-likeness scores; The method (100) of any one of claims 1 to 8.

10. The method (100) comprises: In response to the classification score (5) and / or the occupancy (9) and / or the object-likeness score indicating (160) the presence of an object (O), - obtaining (170) depth information (12) for said image region associated with said object (O); - determining (180) whether the depth information (12) indicates a depth change that may be expected if the object (O) is present; - if the determination is negative, determining (190) that the detection of said object (O) is a false detection; further comprising:

10. The method (100) of claim 9.

11. The method (100) comprises: In response to the classification score (5) and / or the occupancy (9) and / or the object-likeness score indicating (160) the presence of an object (O), Calculating (200) the product (13) of the highest classification score (5) associated with said detected object (O) and said occupancy (9); If the product (13) is below a predetermined threshold (14) (210), determining (220) that the detection of the object (O) is a false detection; further comprising: The method (100) according to claim 9 or 10.

12. The method (100) comprises: calculating (230) an action signal (230a) based at least in part on the classification scores (5) and / or occupancy rates (9) output by the trained machine learning model (1) and / or based at least in part on the detection of an object (O); - operating (240) a vehicle (50), a driver assistance system (51), a robot (60), a quality inspection system (70), a monitoring system (80), and / or a medical imaging system (90) by said operating signal (230a); further comprising: The method (100) of any one of claims 9 to 11.

13. 13. A computer program comprising machine-readable instructions that, when executed by one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the method (100) of any one of claims 1 to 12.

14. A non-transitory machine-readable data carrier and / or download product comprising a computer program according to claim 13.

15. One or more computers and / or computing instances comprising a computer program according to claim 13 and / or comprising a non-transitory machine-readable data carrier and / or download product according to claim 14.

Images