Cooperative-contrastive learning systems and methods
Systems and methods for multi-view cooperative contrastive self-supervised learning, may include receiving a plurality of video sequences, the video sequences comprising a plurality of image frames; applying selected images of a first and second video sequence of the plurality of video sequences to a plurality of different encoders to derive a plurality of embeddings for different views of the selected images of the first and second video sequences; determining distances of the derived plurality of embeddings for the selected images of the first and second video sequences; detecting inconsistencies in the determined distances; and predicting semantics of a future image based on the determined distances.
1. A method for multi-view self-supervised learning, comprising:
receiving a plurality of video sequences, the video sequences comprising a plurality of image frames;
applying selected images of a first and second video sequence of the plurality of video sequences to a plurality of different encoders to derive a plurality of embeddings for different views of the selected images of the first and second video sequences, the plurality of embeddings comprising RGB embeddings, flow embeddings, and KeyPoint embeddings;
determining distances of the derived plurality of embeddings for the selected images of the first and second video sequences;
detecting inconsistencies between distances of the RGB embeddings, distances of the flow embeddings, and distances of the KeyPoint embeddings outside a threshold distance; and
predicting semantics of a future image based on the determined distances.
2. The method of claim 1 , further comprising partitioning the received plurality of video sequences into disjoint blocks; and wherein applying selected images of a first and second video sequence of the plurality of video sequences to a plurality of respective encoders to derive a plurality of embeddings, comprises:
transforming the disjoint blocks into corresponding latent representations for the disjoint blocks; and
generating context representations from the latent representations.
3. The method of claim 2 , wherein predicting future semantics based on the determined distances comprises capturing contextual semantics and frame level semantics to predict a latent state of future images.
4. The method of claim 1 , further comprising using predicted semantics of a future image to predict semantics of an image that is farther into the future than the future image.
5. The method of claim 1 , further comprising computing a noise contrastive estimation loss over at least some of the plurality of embeddings.
6. The method of claim 5 , wherein the noise contrastive estimation loss comprises an entropy loss distinguishing a positive pair of embeddings from negative pairs of embeddings in a video sequence.
7. The method of claim 1 , wherein the plurality of different encoders for the selected images of the first and second video sequences comprise at least two of a flow encoder, an RGB encoder and a keypoint encoder.
8. The method of claim 1 , wherein determining distances comprises determining view-specific distances for the different views, and wherein the method further comprises synchronizing the distances across all views.
9. The method of claim 8 , further comprising enforcing a consistency loss between distances from each of the different views.
10. The method of claim 8 , further comprising deriving discriminative scores based on pairs of embeddings that comprise positive and negative distances.
11. A system for multi-view self-supervised learning, comprising:
a processor; and
a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations, the operations comprising:
receiving a plurality of video sequences, the video sequences comprising a plurality of image frames;
applying selected images of a first and second video sequence of the plurality of video sequences to a plurality of different encoders to derive a plurality of embeddings for different views of the selected images of the first and second video sequences, the plurality of different encoders comprising RGB embeddings, flow embeddings, and KeyPoint embeddings;
determining distances of the derived plurality of embeddings for the selected images of the first and second video sequences;
detecting inconsistencies between distances of RGB embeddings, distances of the flow embeddings, and distances of KeyPoint embeddings outside a threshold distance; and
predicting semantics of a future image based on the determined distances.
12. The system of claim 11 , wherein the operations further comprise partitioning the received plurality of video sequences into disjoint blocks; and wherein applying selected images of a first and second video sequence of the plurality of video sequences to a plurality of respective encoders to derive a plurality of embeddings, comprises:
transforming the disjoint blocks into corresponding latent representations for the disjoint blocks; and
generating context representations from the latent representations.
13. The system of claim 12 , wherein predicting future semantics based on the determined distances comprises capturing contextual semantics and frame level semantics to predict a latent state of future images.
14. The system of claim 11 , wherein the operations further comprise using predicted semantics of a future image to predict semantics of an image that is farther into the future than the future image.
15. The system of claim 11 , wherein the operations further comprise computing a noise contrastive estimation loss over at least some of the plurality of embeddings.
16. The system of claim 15 , wherein the noise contrastive estimation loss comprises an entropy loss distinguishing a positive pair of embeddings from negative pairs of embeddings in a video sequence.
17. The system of claim 11 , wherein the plurality of different encoders for the selected images of the first and second video sequences comprise at least two of a flow encoder, an RGB encoder and a keypoint encoder.
18. The system of claim 11 , wherein determining distances comprises determining view-specific distances for the different views, and wherein the operations further comprise synchronizing the distances across all views.
19. The system of claim 18 , wherein the operations further comprise enforcing a consistency loss between distances from each of the different views.
20. The system of claim 18 , wherein the operations further comprise deriving discriminative scores based on pairs of embeddings that comprise positive and negative distances.