Permutation-Invariant Convolution (PIC) for Identifying Long-Range Activities

Abstract

A method for identifying long-range activities in a video includes segmenting an input video stream to generate a plurality of sets of frames. For each of the sets of frames, the frame having the highest likelihood of including one or more actions from a predefined set of actions is identified, regardless of its order in the set of frames. A global representation of the input stream is generated based on a pooled representation of the identified frames. The long-range activities in the video stream are classified based on the global representation.

Description

[0001] Cross - Reference to Related Applications

[0002] This application claims the benefit of Greek Patent Application No. 20190100517, filed on November 15, 2019, entitled "PERMUTATION INVARIANT CONVOLUTION (PIC) FOR RECOGNIZING LONG - RANGE ACTIVITIES", the disclosure of which is hereby incorporated by reference in its entirety.

[0003] Field of Disclosure

[0004] Aspects of the present disclosure generally relate to temporal modeling, and more particularly to modeling the temporal structure of long - range activities in videos.

[0005] Background

[0006] An artificial neural network, which may include a group of interconnected artificial neurons (e.g., neuron models), may refer to a computing device or a method to be executed by a computing device. Artificial neural networks can be used in various applications and / or devices, such as Internet Protocol (IP) cameras, Internet of Things (IoT) devices, autonomous vehicles, and / or service robots.

[0007] Conventional neural networks use various operations (such as convolution, self - attention, or vector aggregation) to classify short - range actions. However, these operations do not scale to the requirements of long - range activities. There is a need to improve neural networks to recognize long - range activities.

[0008] Summary

[0009] In one aspect of the present disclosure, a method for recognizing long - range activities is provided. The method includes segmenting an input stream to generate a plurality of frame sets. The method further includes identifying, for each frame set from the plurality of frame sets, a frame having the highest likelihood of including a selected action. Additionally, the method includes generating a global representation of the input stream from a pooled representation of the identified frames. Further, the method includes classifying the long - range activity based on the global representation.

[0010] In another aspect of the present disclosure, a device for identifying long-range activities is provided. The device includes a memory and one or more processors coupled to the memory. The (one or more) processors are configured to segment an input stream to generate a plurality of sets of frames. The (one or more) processors are further configured to identify, for each set of frames from the plurality of sets of frames, a frame having the highest likelihood of including one or more actions from a predefined set of actions. Additionally, the (one or more) processors are configured to generate a global representation of the input stream based on a pooled representation of the identified frames. The (one or more) processors are further configured to classify the long-range activities based on the global representation.

[0011] In another aspect of the present disclosure, an apparatus for identifying long-range activities is provided. The apparatus includes means for segmenting an input stream to generate a plurality of sets of frames. The apparatus further includes means for identifying, for each set of frames from the plurality of sets of frames, a frame having the highest likelihood of including one or more actions from a predefined set of actions. Additionally, the apparatus includes means for generating a global representation of the input stream from a pooled representation of the identified frames. Furthermore, the apparatus includes means for classifying the long-range activities based on the global representation.

[0012] In a further aspect of the present disclosure, a non-transitory computer-readable medium is provided. Program code for identifying long-range activities is encoded on the computer-readable medium. The program code is executed by a processor and includes code for segmenting an input stream to generate a plurality of sets of frames. The program code further includes code for identifying, for each set of frames from the plurality of sets of frames, a frame having the highest likelihood of including one or more actions from a predefined set of actions. Additionally, the program code includes code for generating a global representation of the input stream from a pooled representation of the identified frames. Furthermore, the program code includes code for classifying the long-range activities based on the global representation.

[0013] Additional features and advantages of the present disclosure will be described below. Those skilled in the art should appreciate that the present disclosure can be readily used as a basis for modifying or designing other structures for implementing the same purpose as the present disclosure. Those skilled in the art should also recognize that such equivalent constructions do not depart from the teachings of the present disclosure as set forth in the appended claims. The novel features, which are considered to be characteristics of the present disclosure, will be better understood in conjunction with the accompanying drawings when considered in connection with the following description in terms of its organization and method of operation, together with further objects and advantages. It is to be clearly understood, however, that each drawing is provided for the purpose of illustration and description only and is not intended as a definition of the limits of the present disclosure. Brief Description of the Drawings

[0015] The features, nature, and advantages of the present disclosure will become more apparent when the detailed description set forth below is understood in conjunction with the accompanying drawings, in which like reference numerals throughout the drawings consistently identify corresponding parts.

[0016] Figure 1 Illustrates an example implementation of designing a neural network using a system - on - chip (SOC), including a general - purpose processor, in accordance with certain aspects of the present disclosure.

[0017] Figure 2A , 2B and 2C are diagrams illustrating neural networks in accordance with aspects of the present disclosure.

[0018] Figure 2D is a diagram illustrating an exemplary deep convolutional network (DCN) in accordance with aspects of the present disclosure.

[0019] Figure 3 is a block diagram illustrating an exemplary deep convolutional network (DCN) in accordance with aspects of the present disclosure.

[0020] Figure 4 Illustrates multiple example time structures for long - range activities in accordance with aspects of the present disclosure.

[0021] Figure 5 Illustrates an example of a permutation - invariant convolution (PIC) operation in accordance with aspects of the present disclosure.

[0022] Figure 6 Illustrates an example of a permutation - invariant convolution (PIC) layer in accordance with aspects of the present disclosure.

[0023] Figure 7 Illustrates an example of identifying long - range activities in accordance with aspects of the present disclosure.

[0024] Figure 8 Illustrates a flowchart of a method in accordance with aspects of the present disclosure.

[0025] Detailed Description

[0026] The following detailed description, presented in conjunction with the accompanying drawings, is intended as a description of various configurations and is not intended to represent the only configurations in which the described concepts may be practiced. The detailed description includes specific details to provide a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In some instances, well - known structures and components are shown in block diagram form to avoid obscuring such concepts.

[0027] Based on this teaching, those skilled in the art should appreciate that the scope of the present disclosure is intended to cover any aspect of the present disclosure, whether implemented independently of or in combination with any other aspect of the present disclosure. For example, any number of the aspects described can be used to implement an apparatus or practice a method. Additionally, the scope of the present disclosure is intended to cover such apparatuses or methods practiced using other structures, functionalities, or structures and functionalities that complement or are different from the various aspects of the present disclosure described. It should be understood that any aspect of the present disclosure disclosed can be implemented by one or more elements of the claims.

[0028] The term "exemplary" is used herein to mean "serving as an example, instance, or illustration". Any aspect described herein as "exemplary" need not be construed as superior to or better than other aspects.

[0029] Although specific aspects have been described, numerous variations and permutations of these aspects fall within the scope of the present disclosure. While some benefits and advantages of the preferred aspects have been mentioned, the scope of the present disclosure is not intended to be limited to specific benefits, uses, or objectives. Instead, the aspects of the present disclosure are intended to be broadly applicable to different technologies, system configurations, networks, and protocols, some of which are illustrated as examples in the drawings and the following description of the preferred aspects. The detailed description and the drawings merely illustrate the present disclosure and do not limit the present disclosure, the scope of the present disclosure being defined by the appended claims and their equivalent technical solutions.

[0030] In most cases, conventional action recognition systems use convolution, self-attention, or vector aggregation to classify actions depicted in an input stream (e.g., a video). These conventional action recognition systems can accurately identify actions from input streams with a time span less than a threshold. These actions can be referred to as short-range actions. For example, conventional action recognition systems can accurately identify short-range actions such as skiing, boxing, and fencing.

[0031] In some conventional action recognition systems, temporal modeling employs temporal convolution. Temporal modeling relies on a learned kernel W = {w1|i ∈ [1,…,T]}, where T and C are the kernel size and dimension respectively. At the i-th time step, the input features X w = {x1|i ∈ [1,…,T]} are convolved with the kernel W The output feature is The temporal convolution for such conventional action recognition systems is formulated as:

[0032]

[0033] Using such a convolution operation, the kernel W can learn to detect the exact temporal order of the sequence Xw. However, this convolution operation is sensitive to the exact sequential order of Xw. Thus, conventional action recognition systems do not allow for many temporal configurations of unit action sequences that can occur in long-range activities. In other words, the accuracy of conventional action recognition systems degrades when the temporal span of the input stream is greater than a threshold. These actions can be referred to as long-range activities. Long-range activities can include, for example, cooking or making coffee. Long-range activities can last three minutes, five minutes, ten minutes or longer. In some cases, long-range activities last less than three minutes. Additionally, in some aspects, long-range activities can be characterized as having diverse compositions and chaotic temporal orders (meaning difficult to predict).

[0034] Aspects of the present disclosure relate to modeling the temporal structure of long-range activities in video. In one configuration, the temporal structure is modeled by a neural network layer. For simplicity, this layer will be referred to as a permutation invariant convolution (PIC) layer.

[0035] Compared to conventional vector aggregation systems, aspects of the present disclosure use cascaded layers to learn long-range temporal abstractions. Compared to conventional convolution systems, aspects of the present disclosure provide a receptive field that is temporally permutation invariant to features so that weak temporal structures can be modeled. Finally, compared to conventional self-attention systems, aspects of the present disclosure use shared weights to improve the detection of discriminative visual evidence across long videos as well as noisy videos. Thus, compared to conventional action recognition systems, the accuracy of long-range activity recognition is improved by incorporating (a) PIC layer(s) into a neural network.

[0036] Figure 1 An example implementation of a system-on-chip (SOC) 100 in accordance with certain aspects of the present disclosure is illustrated, which may include a central processing unit (CPU) 102 or a multi-core CPU configured for long-range modeling. Variables (e.g., neural signals and synaptic weights), system parameters associated with a computing device (e.g., a neural network with weights), latencies, frequency slot information, and task information may be stored in a memory block associated with a neural processing unit (NPU) 108, a memory block associated with the CPU 102, a memory block associated with a graphics processing unit (GPU) 104, a memory block associated with a digital signal processor (DSP) 106, the memory block 118, or may be distributed across multiple blocks. Instructions executed at the CPU 102 may be loaded from a program memory associated with the CPU 102 or may be loaded from the memory block 118.

[0037] The SOC 100 may also include additional processing blocks customized for specific functions, such as the GPU 104, DSP 106, connectivity block 110 (which may include fifth-generation (5G) connectivity, fourth-generation long-term evolution (4G LTE) connectivity, Wi-Fi connectivity, USB connectivity, Bluetooth connectivity, etc.), and a multimedia processor 112 that can detect and recognize gestures, for example. In one implementation, the NPU is implemented in the CPU, DSP, and / or GPU. The SOC 100 may also include a sensor processor 114, an image signal processor (ISP) 116, and / or a navigation module 120 (which may include a global positioning system).

[0038] The SOC 100 may be based on the ARM instruction set. In one aspect of the present disclosure, the instructions loaded into the general-purpose processor 102 may include code for segmenting an input stream to generate a plurality of sets of frames. The general-purpose processor 102 may also include code for identifying, for each set of frames from the plurality of sets of frames, the frame having the highest likelihood of including a selected action. The general-purpose processor 102 may further include code for generating a global representation of the input stream from the pooled representations of the identified frames. The general-purpose processor 102 may also include code for classifying long-range activities based on the global representation.

[0039] Deep learning architectures can perform object recognition tasks by learning to represent the input at successively higher levels of abstraction in each layer, thereby constructing useful feature representations of the input data. In this way, deep learning addresses the main bottleneck of traditional machine learning. Before the emergence of deep learning, machine learning approaches to object recognition problems might rely heavily on human-engineered features, perhaps combined with shallow classifiers. Shallow classifiers can be two-class linear classifiers, for example, where the weighted sum of the feature vector components can be compared with a threshold to predict which class the input belongs to. Human-engineered features can be templates or kernels customized by engineers with domain expertise for a specific problem domain. In contrast, deep learning architectures can learn to represent features similar to those that human engineers might design, but it learns through training. Additionally, deep networks can learn to represent and recognize new types of features that humans may not have considered yet.

[0040] Deep learning architectures can learn hierarchical features. For example, if visual data is presented to the first layer, the first layer can learn to identify relatively simple features (such as edges) in the input stream. In another example, if auditory data is presented to the first layer, the first layer can learn to identify spectral power in specific frequencies. A second layer that takes the output of the first layer as input can learn to identify feature combinations, such as identifying simple shapes for visual data or sound combinations for auditory data. For example, higher layers can learn to represent complex shapes in visual data or words in auditory data. Even higher layers can learn to identify common visual objects or spoken phrases.

[0041] Deep learning architectures may perform particularly well when applied to problems with a natural hierarchical structure. For example, the classification of motor vehicles can benefit from first learning to identify wheels, windshields, and other features. These features can be combined in different ways at higher levels to identify cars, trucks, and airplanes.

[0042] Neural networks can be designed with various connectivity patterns. In a feedforward network, information is passed from lower layers to higher layers, where each neuron in a given layer communicates to neurons in the higher layer. As described above, hierarchical representations can be built in successive layers of a feedforward network. Neural networks can also have recurrent or feedback (also known as top-down) connections. In recurrent connections, the output from a neuron in a given layer can be communicated to another neuron in the same layer. Recurrent architectures can help identify patterns that span more than one chunk of input data presented sequentially to the neural network. Connections from neurons in a given layer to neurons in lower layers are called feedback (or top-down) connections. Networks with many feedback connections can be beneficial when the identification of high-level concepts can assist in discerning specific low-level features of the input.

[0043] The connections between the layers of a neural network can be fully connected or locally connected. Figure 2A An example of a fully connected neural network 202 is illustrated. In the fully connected neural network 202, a neuron in the first layer can communicate its output to every neuron in the second layer, such that each neuron in the second layer will receive input from every neuron in the first layer. Figure 2BAn example of a locally connected neural network 204 is illustrated. In the locally connected neural network 204, neurons in the first layer can be connected to a limited number of neurons in the second layer. More generally, the locally connected layers of the locally connected neural network 204 can be configured such that each neuron in a layer will have the same or similar connectivity pattern, but their connection strengths can have different values (e.g., 210, 212, 214, and 216). The locally connected connectivity pattern may result in spatially distinct receptive fields in higher layers, since higher layer neurons in a given region can receive inputs that are tuned through training to the properties of a restricted portion of the total input to the network.

[0044] An example of a locally connected neural network is a convolutional neural network. Figure 2C An example of a convolutional neural network 206 is illustrated. The convolutional neural network 206 can be configured such that the connection strengths associated with the inputs for each neuron in the second layer are shared (e.g., 208). Convolutional neural networks can be well-suited for problems where the spatial location of the input is meaningful.

[0045] One type of convolutional neural network is a deep convolutional network (DCN). Figure 2D A detailed example of a DCN 200 designed to recognize visual features from an image 226 input from an image capture device 230, such as an in-vehicle camera, is illustrated. The DCN 200 of the current example can be trained to identify traffic signs and the numbers provided on the traffic signs. Of course, the DCN 200 can be trained for other tasks, such as identifying lane markings or identifying traffic signals.

[0046] The DCN 200 can be trained with supervised learning. During training, an image (such as the image 226 of a speed limit sign) can be presented to the DCN 200, and then a "forward pass" can be computed to produce an output 222. The DCN 200 can include a feature extraction section and a classification section. Upon receiving the image 226, the convolutional layer 232 can apply a convolutional kernel (not shown) to the image 226 to generate a first set of feature maps 218. As an example, the convolutional kernel of the convolutional layer 232 can be a 5x5 kernel that generates 28x28 feature maps. In this example, since four different feature maps are generated in the first set of feature maps 218, four different convolutional kernels are applied to the image 226 at the convolutional layer 232. The convolutional kernel can also be referred to as a filter or a convolutional filter.

[0047] The first set of feature maps 218 can be subsampled by a max pooling layer (not shown) to generate a second set of feature maps 220. The max pooling layer reduces the size of the first set of feature maps 218. That is, the size of the second set of feature maps 220 (such as 14x14) is smaller than the size of the first set of feature maps 218 (such as 28x28). The reduced size provides similar information to subsequent layers while reducing memory consumption. The second set of feature maps 220 can be further convolved via one or more subsequent convolutional layers (not shown) to generate subsequent sets of feature maps (not shown).

[0048] In Figure 2D the example, the second set of feature maps 220 is convolved to generate a first feature vector 224. Additionally, the first feature vector 224 is further convolved to generate a second feature vector 228. Each feature of the second feature vector 228 can include a number corresponding to a possible feature of the image 226 (such as, "logo", "60", and "100"). A softmax function (not shown) can convert the numbers in the second feature vector 228 into probabilities. Thus, the output 222 of the DCN 200 is the probability that the image 226 includes one or more features.

[0049] In this example, the probabilities of "logo" and "60" in the output 222 are higher than the probabilities of other features (such as "30", "40", "50", "70", "80", "90", and "100") of the output 222. Before training, the output 222 produced by the DCN 200 is likely to be incorrect. Thus, the error between the output 222 and the target output can be calculated. The target output is the ground truth of the image 226 (e.g., "logo" and "60"). The weights of the DCN 200 can then be adjusted so that the output 222 of the DCN 200 is more closely aligned with the target output.

[0050] To adjust the weights, the learning algorithm can calculate a gradient vector for the weights. The gradient can indicate the amount by which the error will increase or decrease if the weights are adjusted. At the top layer, the gradient can directly correspond to the value of the weight connecting the activated neurons in the penultimate layer to the neurons in the output layer. In the lower layers, the gradient can depend on the value of the weights and the calculated error gradients of the higher layers. The weights can then be adjusted to reduce the error. This way of adjusting the weights can be referred to as "backpropagation" because it involves a "backward pass" in the neural network.

[0051] In practice, the error gradient of the weights may be computed over a small number of examples, such that the computed gradient approximates the true error gradient. This approximation method may be referred to as stochastic gradient descent. Stochastic gradient descent may be repeated until the error rate achievable by the entire system has stopped decreasing or until the error rate has reached a target level. After learning, the DCN may be presented with a new image (e.g., the speed limit sign of Image 226) and a forward pass through the network may produce an output 222, which may be considered an inference or prediction of the DCN.

[0052] A deep belief network (DBN) is a probabilistic model that includes multiple layers of hidden nodes. The DBN may be used to extract hierarchical representations of a training data set. The DBN may be obtained by stacking multiple layers of restricted Boltzmann machines (RBMs). An RBM is a type of artificial neural network that can learn a probability distribution over an input set. Since an RBM can learn a probability distribution without information about which class each input should be classified into, RBMs are often used in unsupervised learning. Using a hybrid unsupervised and supervised paradigm, the bottom RBM of the DBN may be trained in an unsupervised manner and may be used as a feature extractor, while the top RBM may be trained in a supervised manner (on the joint distribution of the inputs from the previous layer and the target classes) and may be used as a classifier.

[0053] A deep convolutional network (DCN) is a network of convolutional networks that is configured with additional pooling and normalization layers. The DCN has achieved state-of-the-art performance on many tasks. The DCN may be trained using supervised learning, where both the inputs and the output targets are known for many paradigms and are used to modify the weights of the network by using gradient descent.

[0054] The DCN may be a feedforward network. Additionally, as described above, the connections from the neurons in the first layer of the DCN to the groups of neurons in the next higher layer are shared across the neurons in the first layer. The feedforward and shared connections of the DCN may be used for fast processing. The computational burden of the DCN may be much smaller than, for example, that of a neural network of similar size that includes recurrent or feedback connections.

[0055] The processing of each layer of a convolutional network can be considered as a spatially invariant template or basis projection. If the input is first decomposed into multiple channels, such as the red, green, and blue channels of a color image, then a convolutional network trained on that input can be considered three-dimensional, having two spatial dimensions along the axes of the image and a third dimension capturing color information. The output of a convolutional connection can be considered to form a feature map in subsequent layers, where each element in the feature map (e.g., 220) receives input from a certain range of neurons in the previous layer (e.g., feature map 218) and from each of the multiple channels. The values in the feature map can be further processed with a non-linearity (such as rectification, max(0,x)). Values from adjacent neurons can be further pooled (which corresponds to downsampling) and can provide additional local invariance as well as dimensionality reduction. Normalization can also be applied through lateral inhibition between neurons in the feature map, which corresponds to whitening.

[0056] The performance of deep learning architectures can improve as more labeled data points become available or as computing power increases. Modern deep neural networks are routinely trained with thousands of times more computing resources than were available to a typical researcher just fifteen years ago. New architectures and training paradigms can further boost the performance of deep learning. Rectified linear units can reduce a training problem known as vanishing gradients. New training techniques can reduce over-fitting and thus enable larger models to achieve better generalization. Encapsulation techniques can abstract the data within a given receptive field and further improve overall performance.

[0057] Figure 3 is a block diagram illustrating a deep convolutional network 350 in accordance with aspects of the present disclosure. The deep convolutional network 350 can include multiple different types of layers based on connectivity and weight sharing. As Figure 3 shown, the deep convolutional network 350 includes convolutional blocks 354A, 354B. Each of the convolutional blocks 354A, 354B can be configured with a convolutional layer (CONV) 356, a normalization layer (LNorm) 358, and a max pooling layer (MAX POOL) 360.

[0058] The convolutional layer 356 can include one or more convolutional filters that can be applied to input data to generate a feature map. Although only two convolutional blocks 354A, 354B are shown, the present disclosure is not limited thereto, but instead any number of convolutional blocks 354A, 354B can be included in the deep convolutional network 350 according to design preferences. The normalization layer 358 can normalize the output of the convolutional filters. For example, the normalization layer 358 can provide whitening or lateral inhibition. The max pooling layer 360 can provide spatially downsampled aggregation to achieve local invariance as well as dimensionality reduction.

[0059] For example, the parallel filter banks of the deep convolutional network can be loaded onto the CPU 102 or GPU 104 of the SOC 100 to achieve high performance and low power consumption. In an alternative embodiment, the parallel filter banks can be loaded onto the DSP 106 or ISP 116 of the SOC 100. Additionally, the deep convolutional network 350 can access other processing blocks that may be present on the SOC 100, such as the sensor processor 114 and navigation module 120 dedicated to sensors and navigation, respectively.

[0060] The deep convolutional network 350 may also include one or more fully connected layers 362 (FC1 and FC2). The deep convolutional network 350 may further include a logistic regression (LR) layer 364. There are weights (not shown) to be updated between each layer 356, 358, 360, 362, 364 of the deep convolutional network 350. The output of each layer (e.g., 356, 358, 360, 362, 364) can be used as the input to a subsequent layer (e.g., 356, 358, 360, 362, 364) in the deep convolutional network 350 to learn hierarchical feature representations from the input data 352 (e.g., images, audio, video, sensor data, and / or other input data) supplied at the first convolutional block 354A. The output of the deep convolutional network 350 is a classification score 366 for the input data 352. The classification score 366 can be a set of probabilities, where each probability is the probability that the input data includes features from the feature set.

[0061] Long-range human activities can have various characteristics, such as long duration, complex composition, and / or arbitrary temporal order. Coffee preparation is an example of a long-range human activity. In some cases, coffee preparation may take ten minutes from brewing to the final drinking step. Like most long-range activities, coffee preparation consists of multiple short building blocks (e.g., actions), such as "get a cup" and "pour milk". The temporal order of the building blocks can vary based on the individual performing the activity.

[0062] Figure 4 Multiple example temporal structures 400 for the long-range activity of pouring coffee are illustrated in accordance with aspects of the present disclosure. As Figure 4 shown, coffee preparation can include multiple actions. For clarity, Figure 4 the actions of

[0063] Each temporal structure (v1, v2, …, vn) of the video example can be obtained from different input streams. Due to personal preferences, coffee preparation can have many variations. Therefore, the order of actions in each temporal structure may be different. For example, as Figure 4 shown in

[0064] In the first temporal structure v1, the action of "spoon sugar" comes after the action of "pour sugar". Conversely, in the second temporal structure v2, the action of "pour milk" comes after the action of "pour sugar". Figure 4 As shown in

[0065] Long-range activities exhibit complex temporal structures with weaker temporal order. The temporal structure 400 can be described as a partially ordered set with macroscopic and microscopic levels. The information at the macroscopic and microscopic levels can be used to model the temporal structure of long-range activities. At the macroscopic level, long-range activities are subdivided into sets of actions (e.g., unit actions) (the set can also be referred to as a segment). For example, coffee preparation can include three action sets (v = {s1, s2, s3}), where v is the temporal structure and si is the action set. Adjacent actions tend to fall into small sets regardless of the order.

[0066] As an example, for the first temporal structure v1, set one s1 can include "get cup" and "pour coffee". The actions of set one s1 are typically performed at the beginning of the activity. Set two s2 can include "pour sugar", "spoon sugar", and "pour milk". The actions of set two s2 are typically performed in the middle of the activity. Set three s3 can include "stir coffee". The actions of set three s3 are typically performed at the end of the activity.

[0067] In the second temporal structure v2, set one s1 can include "pour coffee" and "get cup". Although the order of set one s1 for the second temporal structure v2 is different from that of set one s1 for the first temporal structure v1, both sets include the same actions that occur at the beginning of the activity. That is, at the microscopic level, each set or segment can include the same combination of actions regardless of the video example of a particular activity. Although the actions in each set may not have a universal order, the actions tend to occur together in the same set regardless of the video.

[0068] As discussed, conventional systems use different time modeling solutions such as convolution, self-attention, and vector aggregation. These solutions can accurately identify activities in short-range videos. For long-range videos, some of the proposed solutions include vector aggregation, long-term features, and long-range convolution. For various reasons, the accuracy of these solutions may be limited. For example, conventional solutions do not learn long-range temporal abstractions. As another example, conventional solutions are not invariant to temporal order. Finally, some conventional solutions do not share weights, and as a result, these solutions cannot detect discriminative concepts.

[0069] In one configuration, a time modeling layer (e.g., PIC) is used to identify long-range activities. The time modeling layer improves the accuracy of long-range activity classification. For simplicity, the time modeling layer may be referred to as PIC. Compared with vector aggregation and self-attention, PIC takes into account local connectivity, such as long-time abstractions learned through cascaded layers. Additionally, compared with conventional convolution, PIC is invariant to temporal permutations within the window of local connectivity.

[0070] Aspects of the present disclosure are not limited to modeling videos (e.g., red-green-blue (RGB) modality). Other modalities such as flow, dynamics, and sound can be used.

[0071] As discussed, the structure of a long-range activity can be defined as a partially ordered set with two levels of abstraction (e.g., macro level and micro level). At the macro level, the entire video v of a long-range activity includes a set of actions (v = {s1, s2, …, sN}) that can be structured and ordered over time. At the micro level, each set includes related actions without a specific order. Each action in the set can be referred to as a one-action.

[0072] The two-level structure of the partially ordered set can be learned using a convolutional method with hierarchical cascades. The bottom layer learns the correlations between actions within each set. The top layer learns the interactions between sets. The convolutional operation is invariant to the temporal order of single actions. For this reason, PIC can replace the convolutional layer (e.g., Figure 3 convolutional layer 356).

[0073] For simplicity, the present disclosure focuses on time modeling. The backbone neural network can focus on spatial modeling. In a conventional system, kernels are learned for convolutional layers. The learned kernels are convolved with a feature window to generate features of the input. However, the learned kernels still depend on the time series of the feature window.

[0074] In contrast, PIC is a permutation-invariant convolution. Figure 5 Illustrates an example of the function of the PIC layer 500 according to aspects of the present disclosure. As Figure 5As shown, the PIC layer 500 receives the feature window Xw as input. The feature window Xw is a window of T frames (x1, x2, …, xT) out of the total number of frames in video 502, where T is greater than or equal to three. Each frame (e.g., x1) may depict one or more actions. The feature window Xw may also be referred to as a local window. For ease of explanation, Figure 5 relates to the activity of pouring coffee. Figure 5 The actions and digital identifiers correspond to Figure 4 the actions and digital identifiers of

[0075] As Figure 5 shown, the first frame x1 has the possibility of depicting the following actions: picking up the cup (1) and pouring coffee (2). The PIC layer 500 uses a pair of linked kernels (referred to as the concept key K (shown as key (K)) and value ), where M is the number of kernels and C is the kernel dimension. The concept key K detects implicit visual concepts in video 502. The similarity matrix is generated by measuring the similarity of the concept key K to the features of each frame of the feature window Xw using the dot product. That is, the similarity matrix s encodes the possibility of implicit concepts in the current feature window Xw. In one configuration, at each layer of the hierarchy, each concept key in K detects one action in the action set.

[0076] As Figure 5 shown, the similarity matrix s (shown as similarity (s)) encodes the first frame x1 and the last frame xT as having the possibility of depicting the activity of picking up the cup (1). In Figure 5 this, the first element 504 of the similarity matrix s has a bold activity identifier (1) to indicate that the first element 504 has a higher possibility of depicting the activity of picking up the cup (1) compared to the second element 506. The bold text of the activity identifier of the elements in the similarity matrix s indicates a greater possibility compared to the non-bold text of the activity identifier of the elements. The possibility can be determined based on the similarity measurement performed using the concept key K.

[0077] The similarities in the similarity matrix s are max-pooled to generate the max-pooled similarity vector The max-pooled similarity vector s’ outlines the frames with the highest possibility of the occurrence of M concepts in the feature window Xw. That is, the max-pooled similarity vector s’ is populated with the frames having the highest (e.g., maximum) possibility of depicting actions. For example, in Figure 5 this, the first element 504 fills the first element of the similarity vector s’ because the first element 504 has a greater possibility of depicting the action (1) of picking up the cup compared to the second element 506 and the third element 508.

[0078] After estimating the pooled similarity vector s', the value kernel V is applied to represent the detected features. In some aspects, the value kernel V is applied to represent only the detected features. The concept key kernel K and the value kernel V decouple concept detection (via the concept key kernel K) from concept representation (via the value kernel V). Decoupling the concept key kernel K from the value kernel V beneficially provides more concept keys for detection and fewer values for representation where M′ << M.

[0079] Dense layer f θ (·) models the correlation between the similarities of the similarity vector s'. The correlation is embedded from a higher dimension to a lower dimension An activation function σ (e.g., sigmoid or rectified linear unit (ReLU)) can be applied to the dense layer f θ (·) to generate the activated similarity (which can also be referred to as the attention vector α). Finally, the dot product of the activated similarity α and the value kernel V is computed to provide the representation T represents the transpose of a vector / matrix.

[0080] Thus, compared to the temporal convolution of a conventional action recognition system (see Equation 1), the permutation invariant convolution (PIC) can be formulated as:

[0081]

[0082] s′ = max row (s) (3)

[0083] α = σ[f θ (s′)] (4)

[0084]

[0085] where ⊙ represents the Hadamard product, and represents the tensor product.

[0086] The PIC layer 500 can be added as a layer to a backbone two - or three - dimensional convolutional neural network (CNN) (e.g., Figure 3 the deep convolutional network 350). In one configuration, the PIC layer 500 uses a residual bottleneck to reduce computation. For example, the dimension C of the feature window can be reduced by a reduction value. As an example, the reduction value can be four, such that the reduced dimension C′ = C / 4. The dimension can be reduced by a first dense layer g φ (·) before the convolution. To enable the residual connection, the input dimension C can be restored by a second dense layer h ψ (·).

[0087] The spatial dimension can be modeled by the backbone CNN. The kernels K, V learned by the PIC layer 500 are shared parameters (e.g., weights) and do not infer or depend on the feature window X w from the feature window X w . The shared kernels can improve the ability of the neural network to detect discriminative visual concepts across the entire long-range activity without conditioning on the feature window X w . This design can provide stability against noisy evidence in the feature window X w .

[0088] In addition, the PIC layer 500 follows temporal locality. In other words, compared to the global window used in self-attention, the PIC layer 500 convolves the features of the local window X w . Because the PIC layer 500 follows temporal locality, the PIC layer 500 can learn multiple levels of abstraction with cascaded layers.

[0089] Figure 6 Illustrates another example of a permutation invariant convolution (PIC) layer 600 according to aspects of the present disclosure. As Figure 6 shown, a feature window X with dimensions T x C w is received at the first dense layer g φ (·). A feature window X with reduced dimensions C’ is received at the PIC layer 600 w . The PIC layer 600 also receives a concept key kernel K and a value kernel V with dimensions M x C’. The PIC layer 600 outputs a representation y with dimensions 1 x C’ to the second dense layer h ψ (·). The second dense layer h ψ (·) restores the input dimension C.

[0090] Figure 7 Illustrates an example of long-range activity in an identification video 700 according to aspects of the present disclosure. As Figure 7 shown, a video 700 of making pancakes is input into a neural network. The video 700 is segmented into feature windows 702 of size N. That is, each feature window 702 includes N frames (x1 to xN). In Figure 7 , N is 5. Of course, N can be greater than or less than 5. In one configuration, N is odd.

[0091] Each feature window 702 is processed by a first permutation invariant convolution (PIC) layer 704. In Figure 7 the example, the first PIC layer 704 identifies a feature from a feature set. For example, the first PIC layer 704 can be specified to identify the frame with the greatest likelihood of depicting a pouring action.

[0092] As discussed, from the similarity matrix, the max pooling function identifies the frame with the highest likelihood of depicting a particular action from a feature window (e.g., 702). In Figure 7 In, the solid line 706a corresponds to the frame with the highest likelihood of depicting a dumping action from each feature window 702. The dashed line 706b corresponds to the frame with a lower likelihood of depicting a dumping action from each feature window 702.

[0093] The max pooling function generates a set of frames 708 from the set of frames identified as having the highest likelihood of depicting a dumping action. For simplicity, Figure 7 illustrates the frame (e.g., 706a) with the highest likelihood of depicting a dumping action. Aspects of the present disclosure are not limited to identifying one concept. Multiple concepts can be identified from the video 700. For example, a local pooling method can be utilized to pool the features of the frames. In local pooling, adjacent features are max pooled. For example, every two adjacent frames are max pooled. In this example, there are eight features in the second layer regardless of the activation values.

[0094] The second PIC layer 710 is designated to identify fine-grained concepts. For example, the dumping activity is not related to activities such as "making coffee", "making tea", and "making pancakes". Thus, the second PIC layer 710 identifies the activity corresponding to the dumping action.

[0095] The set of frames 708 is segmented into feature windows 712 of size N. The second PIC layer 710 identifies the frame with the highest likelihood of depicting an action from an action set. The frame with the highest likelihood is identified and pooled in the global pooling function to generate a final output 714, which can be a global representation of the activity observed in the video stream. For example, in Figure 7 In, the second PIC layer 710 can identify the frame with the highest likelihood of depicting a cooking activity. In Figure 7 In, the solid line 716 corresponds to the frame with the highest likelihood of depicting a dumping action from each feature window 712. The final output 714 can then be classified as a pancake-making action. For example, in one configuration, a two-hidden-layer multi-layer perceptron can provide the classification.

[0096] As discussed, according to aspects of the neural network, the model can learn the unordered temporal representation of atomic actions in a video (e.g., 700). Thus, if the input segments of the video (e.g., 700) are reordered (e.g., randomly shuffled), the accuracy of the model may not be compromised. The model can classify action videos with relatively long time spans (e.g., greater than or equal to several minutes). Finally, the model is invariant to horizontal flips.

[0097] Figure 8 Illustrates a flowchart 800 of a method 800 according to aspects of the present disclosure. AsFigure 8 As shown in Figure 8 , at block 802, a neural network segments an input stream to generate multiple sets of frames. The input stream can be a video of a long-range activity. At block 804, the neural network identifies, for each set of frames from the multiple sets of frames, the frame having the highest likelihood of including one or more actions from a predefined set of actions.

[0098] The frame can be identified by generating a similarity matrix from the dot product of the features of the set of frames and a first kernel. The similarities in the similarity matrix can be max-pooled to identify the frame having the highest likelihood. At block 806, the neural network generates a global representation of the input stream from the pooled representations of the identified frames. The representation of the identified frames can be generated by: generating an attention vector from the similarity matrix; and generating the representation of the identified frames based on the dot product of the attention vector and a second kernel. In one configuration, the first kernel and the second kernel are linked.

[0099] The global representation can be generated from the dot product of the representations of the identified frames. At block 808, the neural network classifies the long-range activity based on the global representation. The permutation-invariant convolutional layer of the neural network can identify the frame having the highest likelihood in the segmentation of the frames and generate a representation. The neural network can include multiple cascaded permutation-invariant convolutional layers.

[0100] In some aspects, method 800 can be performed by SOC 100 ( Figure 1 ). That is, by way of example and not limitation, each element of method 800 can be performed by SOC 100 or one or more processors (e.g., CPU 102) and / or other included components.

[0101] Implementation examples are described in the following numbered clauses:

[0102] 1. A method, comprising:

[0103] Segmenting an input stream to generate multiple sets of frames;

[0104] Identifying, for each set of frames from the multiple sets of frames, the frame having the highest likelihood of including one or more actions from a predefined set of actions;

[0105] Generating a global representation of the input stream from the pooled representations of the identified frames; and

[0106] Classifying a long-range activity based on the global representation.

[0107] 2. The method of clause 1, wherein identifying the frame includes generating a similarity matrix from the dot product of the features of the set of frames and a first kernel.

[0108] 3. The method of clause 2, further comprising: max-pooling the similarities in the similarity matrix to identify the frame having the highest likelihood.

[0109] 4. The method as in clause 3, further comprising:

[0110] Generating an attention vector from the similarity matrix; and

[0111] Generating a global representation of the identified frame based on the dot product of the attention vector and the second kernel.

[0112] 5. The method as in clause 4, wherein the first kernel and the second kernel are linked.

[0113] 6. The method as in clause 1, wherein the global representation is based on the dot product of the pooled representations of the identified frames.

[0114] 7. The method as in clause 1, wherein the identification and generation are performed at a permutation invariant convolutional layer of a neural network.

[0115] 8. The method as in any one of clauses 1-7, wherein the neural network comprises a plurality of cascaded permutation invariant convolutional layers.

[0116] 9. An apparatus, comprising:

[0117] A memory; and

[0118] At least one processor coupled to the memory, the at least one processor being configured to:

[0119] Segment an input stream to generate a plurality of sets of frames;

[0120] For each set of frames from the plurality of sets of frames, identify a frame having the highest likelihood of including one or more actions from a predefined set of actions;

[0121] Generate a global representation of the input stream from the pooled representations of the identified frames; and

[0122] Classify a long-range activity based on the global representation.

[0123] 10. The apparatus as in clause 9, wherein the at least one processor is configured to identify the frame by generating a similarity matrix from the dot product of the features of the set of frames and the first kernel.

[0124] 11. The apparatus as in clause 10, wherein the at least one processor is further configured to perform max pooling on the similarities in the similarity matrix to identify the frame having the highest likelihood.

[0125] 12. The apparatus as in clause 11, wherein the at least one processor is further configured to generate a global representation by:

[0126] Generating an attention vector from the similarity matrix; and

[0127] Generate a global representation of the identified frame based on the dot product of the attention vector and the second kernel.

[0128] 13. The apparatus of clause 12, wherein the first kernel and the second kernel are linked.

[0129] 14. The apparatus of clause 9, wherein the at least one processor is further configured to generate the global representation based on the dot product of the pooled representation of the identified frame.

[0130] 15. The apparatus of clause 9, wherein the at least one processor is further configured to identify and generate at a permutation-invariant convolutional layer of a neural network.

[0131] 16. The apparatus of any one of clauses 9-15, wherein the neural network includes a plurality of cascaded permutation-invariant convolutional layers.

[0132] 17. An apparatus, comprising:

[0133] means for segmenting an input stream to generate a plurality of sets of frames;

[0134] means for identifying, for each set of frames from the plurality of sets of frames, a frame having the highest likelihood of including one or more actions from a predefined set of actions;

[0135] means for generating a global representation of the input stream from the pooled representation of the identified frames; and

[0136] means for classifying a long-range activity based on the global representation.

[0137] 18. The apparatus of clause 17, further comprising means for generating a similarity matrix from the dot product of the features of the set of frames and the first kernel.

[0138] 19. The apparatus of clause 18, further comprising means for performing max pooling on the similarities in the similarity matrix to identify the frame having the highest likelihood.

[0139] 20. The apparatus of clause 19, further comprising:

[0140] means for generating an attention vector from the similarity matrix; and

[0141] means for generating a global representation of the identified frame based on the dot product of the attention vector and the second kernel.

[0142] 21. The apparatus of clause 20, wherein the first kernel and the second kernel are linked.

[0143] 22. An apparatus as in any of clauses 17 - 21, wherein the global representation is a dot product of a pooled representation of the identified frames.

[0144] 23. A non - transient computer - readable medium having program code recorded thereon, the program code being executed by a processor and including:

[0145] Program code for segmenting an input stream to generate a plurality of sets of frames;

[0146] Program code for identifying, for each set of frames from the plurality of sets of frames, a frame having the highest likelihood of including one or more actions from a predefined set of actions;

[0147] Program code for generating a global representation of the input stream from a pooled representation of the identified frames; and

[0148] Program code for classifying long - range activities based on the global representation.

[0149] 24. The non - transient computer - readable medium as in clause 23, further including program code for identifying frames by generating a similarity matrix from a dot product of features of a set of frames and a first kernel.

[0150] 25. The non - transient computer - readable medium as in clause 24, further including: program code for performing max - pooling on the similarities in the similarity matrix to identify the frame having the highest likelihood.

[0151] 26. The non - transient computer - readable medium as in clause 25, further including program code for:

[0152] Generating an attention vector from the similarity matrix; and

[0153] Generating a global representation of the identified frames based on a dot product of the attention vector and a second kernel.

[0154] 27. The non - transient computer - readable medium as in clause 26, wherein the first kernel and the second kernel are linked.

[0155] 28. The non - transient computer - readable medium as in clause 23, further including program code for generating the global representation based on a dot product of a pooled representation of the identified frames.

[0156] 29. The non - transient computer - readable medium as in clause 23, further including program code for performing identification and generation at a permutation - invariant convolutional layer of a neural network.

[0157] 30. The non - transient computer - readable medium as in any of clauses 23 - 29, wherein the neural network includes a plurality of cascaded permutation - invariant convolutional layers.

[0158] The various operations of the methods described above can be performed by any suitable device capable of performing the corresponding functions. These devices can include various hardware and / or software components and / or modules, including but not limited to circuits, application specific integrated circuits (ASICs), or processors. Generally, where operations are illustrated in the figures, those operations can have corresponding paired apparatus plus function components with similar numbers.

[0159] As used, the term "determine" encompasses a variety of actions. For example, "determine" can include computing, calculating, processing, deriving, researching, looking up (e.g., looking up in a table, database, or other data structure), ascertaining, and the like. Additionally, "determine" can include receiving (e.g., receiving information), accessing (e.g., accessing data in a memory), and similar actions. Further, "determine" can include parsing, selecting, choosing, establishing, and the like.

[0160] As used, the phrase that recites "at least one of" a list of items refers to any combination of those items, including a single member. By way of example, "at least one of a, b, or c" is intended to cover: a, b, c, a - b, a - c, b - c, and a - b - c.

[0161] The various illustrative logical blocks, modules, and circuits described in connection with the present disclosure can be implemented or performed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array signal (FPGA) or other programmable logic device (PLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the described functions. The general purpose processor can be a microprocessor, but in the alternative, the processor can be any commercially available processor, controller, microcontroller, or state machine. The processor can also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0162] The steps of the methods or algorithms described in connection with the present disclosure may be implemented directly in hardware, in a software module executed by a processor, or in a combination of both. The software modules may reside in any form of storage medium known in the art. Some examples of storage media that may be used include random access memory (RAM), read-only memory (ROM), flash memory, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, CD-ROMs, and the like. The software modules may include a single instruction, or many instructions, and may be distributed over several different code segments, distributed among different programs, and across multiple storage media. The storage medium may be coupled to the processor such that the processor can read from, and write to, the storage medium. In an alternative, the storage medium may be integrated into the processor.

[0163] The disclosed methods include one or more steps or acts for achieving the described methods. These method steps and / or acts may be interchanged with one another without departing from the scope of the claims. In other words, unless a specific order of the steps or acts is specified, the order and / or use of the specific steps and / or acts may be altered without departing from the scope of the claims.

[0164] The described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in hardware, an example hardware configuration may include a processing system in a device. The processing system may be implemented with a bus architecture. Depending on the particular application and overall design constraints of the processing system, the bus may include any number of interconnecting buses and bridges. The bus may link together various circuits including a processor, a machine-readable medium, and a bus interface. The bus interface may be used to connect, among other things, a network adapter to the processing system via the bus. The network adapter may be used to implement signal processing functions. For some aspects, a user interface (e.g., keypad, display, mouse, joystick, etc.) may also be connected to the bus. The bus may also link various other circuits such as a timing source, peripherals, voltage regulators, power management circuits, and similar circuits, which are well known in the art and will not be described further.

[0165] The processor may be responsible for managing the bus and general processing, including executing software stored on a machine-readable medium. The processor may be implemented with one or more general-purpose and / or special-purpose processors. Examples include microprocessors, microcontrollers, DSP processors, and other circuitry capable of executing software. Software should be construed broadly to mean instructions, data, or any combination thereof, whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise. By way of example, the machine-readable medium may include random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, magnetic disks, optical disks, hard drives, or any other suitable storage medium, or any combination thereof. The machine-readable medium may be embodied in a computer program product. The computer program product may include packaging materials.

[0166] In a hardware implementation, the machine-readable medium may be a part of the processing system separate from the processor. However, as will be readily appreciated by those skilled in the art, the machine-readable medium or any part thereof may be external to the processing system. By way of example, the machine-readable medium may include transmission lines, carrier waves modulated by data, and / or computer products separate from the device, all of which may be accessed by the processor via a bus interface. Alternatively or additionally, the machine-readable medium or any part thereof may be integrated into the processor, such as may be the case with a cache and / or a general register file. Although the various components discussed may be described as having a particular location, such as local components, they may also be configured in various ways, such as some components being configured as part of a distributed computing system.

[0167] The processing system may be configured as a general-purpose processing system having one or more microprocessors providing processor functionality and an external memory providing at least a portion of the machine-readable medium, all linked together via an external bus architecture with other support circuitry. Alternatively, the processing system may include one or more neuromorphic processors for implementing the described neuron models and nervous system models. As another alternative, the processing system may be implemented with an application-specific integrated circuit (ASIC) having a processor, bus interface, user interface, support circuitry, and at least a portion of the machine-readable medium integrated on a single chip, or with one or more field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), controllers, state machines, gated logic, discrete hardware components, or any other suitable circuitry, or any combination of circuits capable of performing the various functions described throughout this disclosure. Depending on the particular application and the overall design constraints imposed on the system, those skilled in the art will recognize how best to implement the functionality described with respect to the processing system.

[0168] A machine-readable medium may include several software modules. These software modules include instructions that cause a processing system to perform various functions when executed by a processor. These software modules may include a transmission module and a reception module. Each software module may reside in a single storage device or be distributed across multiple storage devices. As an example, when a triggering event occurs, the software module may be loaded from a hard drive into RAM. During the execution of the software module, the processor may load some instructions into the cache to improve access speed. One or more cache lines may then be loaded into the general register file for the processor to execute. When referring to the functionality of the software module hereinafter, it will be understood that such functionality is implemented by the processor when the processor executes instructions from the software module. In addition, it should be appreciated that aspects of the present disclosure result in improvements to the functionality of a processor, computer, machine, or other system implementing such aspects.

[0169] If implemented in software, the functions may be stored on or transmitted via a computer-readable medium as one or more instructions or code. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. The storage media may be any available media that can be accessed by a computer. By way of example and not limitation, such computer-readable media may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Additionally, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a web site, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technology such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technology such as infrared, radio, and microwave is included in the definition of the medium. As used herein, disk and disc include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk, and Blu-ray disc, where disk often magnetically reproduces data, while disc optically reproduces data with a laser. Thus, in some aspects, computer-readable media may include non-transitory computer-readable media (e.g., tangible media). Additionally, for other aspects, computer-readable media may include transitory computer-readable media (e.g., signals). Combinations of the above should also be included within the scope of computer-readable media.

[0170] Accordingly, some aspects may include a computer program product for performing the given operations. For example, such a computer program product may include a computer-readable medium having instructions stored (and / or encoded) thereon that are executable by one or more processors to perform the described operations. For some aspects, the computer program product may include packaging material.

[0171] In addition, it should be appreciated that modules and / or other suitable means for performing the described methods and techniques may be downloaded and / or otherwise obtained by a user terminal and / or a base station where applicable. For example, such devices can be coupled to a server to facilitate the transfer of means for performing the described methods. Alternatively, the various methods can be provided via a storage device (e.g., RAM, ROM, a physical storage medium such as a compact disc (CD) or a floppy disk, etc.) such that once the storage device is coupled to or provided to the user terminal and / or the base station, the device can obtain the various methods. In addition, any other suitable technique for providing the described methods and techniques to a device can be utilized.

[0172] It will be understood that the claims are not limited to the exact configurations and components described above. Various changes, substitutions, and modifications can be made in the layout, operation, and details of the methods and apparatuses described above without departing from the scope of the claims.

Claims

1. A method for identifying long-range activities, comprising: Segmenting an input stream to generate a plurality of frame sets; Identifying, by a first permutation invariant convolution (PIC) layer and a first max pooling layer, for each frame set from the plurality of frame sets, a first set of frames having the highest likelihood of including one or more actions from a predefined set of actions; Generating, by the first max pooling layer, a frame set from the first set of frames; Segmenting the frame set into feature windows; Identifying, by a second PIC layer and a second max pooling layer, for each feature window among the feature windows, a second set of frames having the highest likelihood of including one or more actions from a predefined set of actions; Generating a global representation of the input stream from the second set of frames; and Classifying the long-range activities based on the global representation, wherein the first PIC layer and the second PIC layer are invariant to temporal permutations within a window of local connectivity.

2. The method according to claim 1, wherein identifying the frames comprises generating a similarity matrix from the dot product of the features of the frame set and a first kernel.

3. The method according to claim 2, further comprising: Performing max pooling on the similarities in the similarity matrix to identify the frames having the highest likelihood.

4. The method according to claim 3, further comprising: Generating an attention vector from the similarity matrix; and Generating the global representation based on the dot product of the attention vector and a second kernel.

5. The method according to claim 4, wherein the first kernel and the second kernel are linked.

6. The method according to claim 1, wherein the global representation is based on the dot product of the pooled representations of the identified frames.

7. An apparatus for identifying long-range activities, comprising: A memory; and At least one processor coupled to the memory, the at least one processor being configured to: Segment an input stream to generate a plurality of frame sets; Identifying, by a first permutation invariant convolution (PIC) layer and a first max pooling layer, for each frame set from the plurality of frame sets, a first set of frames having the highest likelihood of including one or more actions from a predefined set of actions; Generating, by the first max pooling layer, a frame set from the first set of frames; Segmenting the frame set into feature windows; Identifying, by a second PIC layer and a second max pooling layer, for each feature window among the feature windows, a second set of frames having the highest likelihood of including one or more actions from a predefined set of actions; Generating a global representation of the input stream from the second set of frames; and Classifying the long-range activities based on the global representation, wherein the first PIC layer and the second PIC layer are invariant to temporal permutations within a window of local connectivity.

8. The apparatus according to claim 7, wherein the at least one processor is configured to identify the frames by generating a similarity matrix from the dot product of the features of the frame set and a first kernel.

9. The apparatus according to claim 8, wherein the at least one processor is further configured to perform max pooling on the similarities in the similarity matrix to identify the frames having the highest likelihood.

10. The apparatus according to claim 9, wherein the at least one processor is further configured to: Generate an attention vector from the similarity matrix; and Generate the global representation based on the dot product of the attention vector and a second kernel.

11. The apparatus according to claim 10, wherein the first kernel and the second kernel are linked.

12. The apparatus according to claim 7, wherein the at least one processor is further configured to generate the global representation based on the dot product of the pooled representations of the identified frames.

13. A device for identifying long-range activities, comprising:[[]] means for segmenting an input stream to generate a plurality of sets of frames; means for identifying, by a first permutation invariant convolutional (PIC) layer and a first max pooling layer, for each set of frames from the plurality of sets of frames, a first set of frames having the highest likelihood of including one or more actions from a predefined set of actions; means for generating, by the first max pooling layer, a set of frames from the first set of frames; means for segmenting the set of frames into feature windows; means for identifying, by a second PIC layer and a second max pooling layer, for each feature window in the feature windows, a second set of frames having the highest likelihood of including one or more actions from a predefined set of actions; means for generating a global representation of the input stream from the second set of frames; and means for classifying the long-range activity based on the global representation, wherein the first PIC layer and the second PIC layer are invariant to temporal permutations within a window of local connectivity.

14. The device according to claim 13, further comprising means for generating a similarity matrix from the dot product of the features of the set of frames and a first kernel.

15. The device according to claim 14, further comprising means for performing max pooling on the similarities in the similarity matrix to identify the frames having the highest likelihood.

16. The device according to claim 15, further comprising:[[]] means for generating an attention vector from the similarity matrix; and means for generating the global representation based on the dot product of the attention vector and a second kernel.

17. The device according to claim 16, wherein the first kernel and the second kernel are linked.

18. The device according to claim 13, wherein the global representation is based on the dot product of the pooled representations of the identified frames.

19. A non-transitory computer-readable medium having program code recorded thereon, the program code being executed by a processor and comprising:[[]] program code for segmenting an input stream to generate a plurality of sets of frames; program code for identifying, by a first PIC layer and a first max pooling layer, for each set of frames from the plurality of sets of frames, a first set of frames having the highest likelihood of including one or more actions from a predefined set of actions; program code for generating, by the first max pooling layer, a set of frames from the first set of frames; program code for segmenting the set of frames into feature windows; program code for identifying, by a second PIC layer and a second max pooling layer, for each feature window in the feature windows, a second set of frames having the highest likelihood of including one or more actions from a predefined set of actions; Program code for generating a global representation of the input stream from the second set of frames; and program code for classifying long-range activities based on the global representation, wherein the first PIC layer and the second PIC layer are invariant to temporal permutations within a window of local connectivity.

20. The non-transitory computer-readable medium of claim 19, further comprising program code for identifying the frames by generating a similarity matrix from the dot product of the features of the set of frames and a first kernel.

21. The non-transitory computer-readable medium of claim 20, further comprising program code for performing max pooling on the similarities in the similarity matrix to identify the frames with the highest likelihood.

22. The non-transitory computer-readable medium of claim 21, further comprising program code for: program code for generating an attention vector from the similarity matrix; and program code for generating the global representation based on the dot product of the attention vector and a second kernel.

23. The non-transitory computer-readable medium of claim 22, wherein the first kernel and the second kernel are linked.

24. The non-transitory computer-readable medium of claim 19, further comprising program code for generating the global representation based on the dot product of the pooled representations of the identified frames.

Images