Budget-aware method for detecting activity in video
Detection of activity in video content, and more particularly detecting in video start and end frames inclusive of an activity and a classification for the activity, is fundamental for video analytics including categorizing, searching, indexing, segmentation, and retrieval of videos. Existing activity detection processes rely on a large set of features and classifiers that exhaustively run over every time step of a video at multiple temporal scales, or as a small improvement computationally propose segments of the video on which to perform classification. These existing activity detection processes, however, are computationally expensive, particularly when trying to achieve activity detection accuracy, and moreover are not configurable for any particular time or computation budget. The present disclosure provides a time and/or computation budget-aware method for detecting activity in video that relies on a recurrent neural network implementing a learned policy.
1. A method for detecting activity in video, comprising:
learning, through a reinforcement learning algorithm, a policy to sequentially select frames of a video and classify activity within the video according to a plurality of predefined activity classifications;
using, by a recurrent neural network, the policy to detect and classify activity in the video over a sequence of steps, wherein a number of steps in the sequence of steps is determined according to a given time and/or computation budget, and wherein the policy, at each step in the sequence of steps, is configured to:
receive a location of a frame to observe during the step,
select a subset of frames of the video within a neighborhood of the location of the frame, and
predict, based on the subset of frames and a history of any prior observed frames, a temporal location of a segment in the video associated with an activity, one or more of the plurality of predefined activity classifications associated with the segment, and a next location for a different frame in the video to select for a next step in the sequence of steps.
2. The method of claim 1 , wherein the policy is parameterized by θ, and the learning is to optimize θ by minimizing a loss associated with a set of estimated segments in the video.
3. The method of claim 2 , wherein the loss refers to at least one of a classification loss, a localization loss, and a retrieval loss.
4. The method of claim 1 , wherein the policy is learned using a partially observable Markov decision process (POMDP).
5. The method of claim 4 , wherein the policy is learned using approximation of gradients for an objective function using a recurrent policy gradient approach.
6. The method of claim 5 , wherein parameters of the policy are learned with stochastic gradient descent.
7. The method of claim 1 , wherein the reinforcement learning algorithm rewards the policy as a function of a change in error.
8. The method of claim 7 , wherein the policy earns a reward equal to any decrease in a temporal segmentation error achieved by selecting a particular frame, and pays a penalty when the temporal segmentation error increases.
9. The method of claim 1 , wherein the plurality of predefined activity classifications includes:
a first set of activity classifications for different foreground activities, and
a second activity classification for background activity.
10. The system of claim 2 , wherein the loss refers to at least one of a classification loss, a localization loss, and a retrieval loss.
11. A system for detecting activity in video, comprising:
a computer processor executing a reinforcement learning algorithm to learn a policy to sequentially select frames of a video and classify activity within the video according to a plurality of predefined activity classifications; and
a recurrent neural network that uses the policy to detect and classify activity in the video over a sequence of steps, wherein a number of steps in the sequence of steps is determined according to a given time and/or computation budget, and wherein the policy, at each step in the sequence of steps, is configured to:
receive a location of a frame to observe during the step,
select a subset of frames of the video within the neighborhood of the location of the frame, and
predict, based on the subset of frames and a history of any prior observed frames, a temporal location of a segment in the video associated with an activity, one or more of the plurality of predefined activity classifications associated with the segment, and a next location for a different frame in the video to select for a next step in the sequence of steps.
12. The system of claim 11 , wherein the policy is parameterized by θ, and the goal of the policy learning is to optimize θ by minimizing a loss associated with a set of estimated segments in the video.
13. The system of claim 11 , wherein the policy is learned using a partially observable Markov decision process (POMDP).
14. The system of claim 13 , wherein the policy is learned using approximation of gradients for an objective function using a recurrent policy gradient approach.
15. The system of claim 14 , wherein parameters of the policy are learned with stochastic gradient descent.
16. The system of claim 11 , wherein the reinforcement learning algorithm rewards the policy as a function of a change in error.
17. The system of claim 16 , wherein the policy earns a reward equal to any decrease in a temporal segmentation error achieved by selecting a particular frame, and pays a penalty when the temporal segmentation error increases.
18. The system of claim 11 , wherein the plurality of predefined activity classifications includes:
a first set of activity classifications for different foreground activities, and
a second activity classification for background activity.
19. A non-transitory computer-readable media storing computer instructions for detecting activity in video that, when executed by one or more processors, cause the one or more processors to perform the method comprising:
learning, through a reinforcement learning algorithm, a policy to sequentially select frames of a video and classify activity within the video according to a plurality of predefined activity classifications;
using, by a recurrent neural network, the policy to detect and classify activity in the video over a sequence of steps, wherein a number of steps in the sequence of steps is determined according to a given time and/or computation budget, and wherein the policy, at each step in the sequence of steps, is configured to:
receive a location of a frame to observe during the step,
select a subset of frames of the video within a neighborhood of the location of the frame, and
predict, based on the subset of frames and a history of any prior observed frames, a temporal location of a segment in the video associated with an activity, one or more of the plurality of predefined activity classifications associated with the segment, and a next location for a different frame in the video to select for a next step in the sequence of steps.
20. The non-transitory computer-readable media of claim 19 , wherein the policy is parameterized by θ, and the goal of the policy learning is to optimize θ by minimizing a loss associated with a set of estimated segments in the video.
21. The non-transitory computer-readable media of claim 19 , wherein the plurality of predefined activity classifications includes:
a first set of activity classifications for different foreground activities, and
a second activity classification for background activity.