Self-supervised cross-video temporal difference learning for unsupervised domain adaptation

1. A computer-implemented method for Cross Video Temporal Difference (CVTD) learning for unsupervised domain adaptation, comprising:

adapting a source domain video to a target domain video using a CVTD loss, wherein the source domain video is annotated, and the target domain video is unannotated,

wherein the CVTD loss is computed by

quantizing clips derived from the source domain video and the target domain video by dividing the source domain video into a plurality of source domain clips and the target domain video into a plurality of target domain clips;

sampling two clips from each of the plurality of source domain clips and the plurality of target domain clips to obtain four sampled clips comprising a first sampled source domain clip, a second sampled source domain clip, a first sampled target domain clip, and a second sampled target domain clip; and

computing, by a clip encoder convolutional neural network, the CVTD loss as |(second sampled source domain clip−first sampled source domain clip)−(second sampled target domain clip−first sampled target domain clip)|.

2. The computer-implemented method of claim 1 , further comprising concatenating, by the clip encoder convolutional neural network, features of the four sampled clips to obtain a set of concatenated features used to calculate the CVTD loss.

3. The computer-implemented method of claim 1 , wherein the CVTD loss is an L 2 loss between a predicted CVTD value versus a true CVTD value for the four sampled clips.

4. The computer-implemented method of claim 1 , wherein the source domain video is adapted to the target domain video further using a source classification loss that inputs the plurality of source domain clips and optimized predictions of classes thereon.

5. The computer-implemented method of claim 4 , wherein the source classification loss is a cross entropy loss.

6. The computer-implemented method of claim 4 , wherein the source classification loss comprises per class sigmoid losses.

7. The computer-implemented method of claim 1 , wherein the source domain video is adapted to the target domain video further using a domain adversarial loss.

8. The computer-implemented method of claim 1 , wherein the CVTD loss is applied to learn representations capable of predicting temporal differences between the plurality of source domain clips and the plurality of target domain clips.

9. The computer-implemented method of claim 1 , wherein said sampling step samples the two clips from each of the plurality of source domain clips and the plurality of target domain clips to obtain the four sampled clips using importance-based sampling that excludes background only images.

10. The computer-implemented method of claim 9 , wherein during training, a relative importance of the pluralities of source and target domain clips are updated based on a cross-entropy of the plurality of source domain videos and an entropy of predictions of the plurality of target domain videos.

11. A computer program product for Cross Video Temporal Difference (CVTD) learning for unsupervised domain adaptation, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:

adapting a source domain video to a target domain video using a CVTD loss, wherein the source domain video is annotated, and the target domain video is unannotated,

wherein the CVTD loss is computed by

quantizing clips derived from the source domain video and the target domain video by dividing the source domain video into a plurality of source domain clips and the target domain video into a plurality of target domain clips;

sampling two clips from each of the plurality of source domain clips and the plurality of target domain clips to obtain four sampled clips comprising a first sampled source domain clip, a second sampled source domain clip, a first sampled target domain clip, and a second sampled target domain clip; and

computing, by a clip encoder convolutional neural network, the CVTD loss as |(second sampled source domain clip−first sampled source domain clip)−(second sampled target domain clip−first sampled target domain clip)|.

12. The computer program product of claim 11 , wherein the method further comprises concatenating, by the clip encoder convolutional neural network, features of the four sampled clips to obtain a set of concatenated features used to calculate the CVTD loss.

13. The computer program product of claim 11 , wherein the CVTD loss is an L 2 loss between a predicted CVTD value versus a true CVTD value for the four sampled clips.

14. The computer program product of claim 11 , wherein the source domain video is adapted to the target domain video further using a source classification loss that inputs the plurality of source domain clips and optimized predictions of classes thereon.

15. The computer program product of claim 14 , wherein the source classification loss is a cross entropy loss.

16. The computer program product of claim 14 , wherein the source classification loss comprises per class sigmoid losses.

17. The computer program product of claim 11 , wherein the source domain video is adapted to the target domain video further using a domain adversarial loss.

18. The computer program product of claim 11 , wherein the CVTD loss is applied to learn representations capable of predicting temporal differences between the plurality of source domain clips and the plurality of target domain clips.

19. The computer program product of claim 11 , wherein said sampling step samples the two clips from each of the plurality of source domain clips and the plurality of target domain clips to obtain the four sampled clips using importance-based sampling that excludes background only images.

20. A computer processing system for Cross Video Temporal Difference (CVTD) learning for unsupervised domain adaptation, comprising:

a memory device for storing program code; and

a hardware processor operatively coupled to the memory device for running the program code to adapt a source domain video to a target domain video using a CVTD loss, wherein the source domain video is annotated, and the target domain video is unannotated,

wherein the hardware processor computes the CVTD loss by

quantizing clips derived from the source domain video and the target domain video by dividing the source domain video into a plurality of source domain clips and the target domain video into a plurality of target domain clips;

sampling two clips from each of the plurality of source domain clips and the plurality of target domain clips to obtain four sampled clips comprising a first sampled source domain clip, a second sampled source domain clip, a first sampled target domain clip, and a second sampled target domain clip; and

computing, using a clip encoder convolutional neural network, the CVTD loss as |(second sampled source domain clip−first sampled source domain clip)−(second sampled target domain clip−first sampled target domain clip)|.