Video recognition using multiple modalities

1. A method, comprising:

classifying a video stream using a plurality of classifier networks of a multi-modality video recognition system, wherein the plurality of classifier networks are trained by;

generating a plurality of video streams of a sequence of object motion using a plurality of modalities,

training a plurality of classifier networks, each of the plurality of classifier networks trained with a different one of the plurality of video streams, wherein each of the plurality of different classifier networks includes multiple intermediate layers,

determining correlation matrices of related intermediate layers of each of the plurality of the different classifier networks,

aligning the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks by minimizing a distance between the correlation matrices, and

improving one of the plurality of classifier network's representation of a specific frame in its feature maps using the correlation matrices of the related intermediate layers of another of the plurality of classifier networks.

2. The method of claim 1 , wherein aligning the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks further comprises determining a spatiotemporal semantic alignment (SSA) loss between the related intermediate layers.

3. The method of claim 2 , wherein determining a spatiotemporal semantic alignment (SSA) loss between the related intermediate layers further comprises:

generating an adaptive regularization parameter; and

regularizing the spatiotemporal semantic alignment loss using the adaptive regularization parameter.

4. The method of claim 3 , wherein generating the adaptive regularization parameter further comprising generating difference between classification losses of the plurality of the different classifier networks and generating the adaptive regularization parameter using the difference between classification losses.

5. The method of claim 3 , wherein generating the adaptive regularization parameter further comprises setting the adaptive regularization parameter to zero to avoid negative knowledge transfer.

6. The method of claim 1 , wherein the plurality of modalities includes two or more of RGB video modality, depth video modality, infrared video modality, and heat video modality.

7. The method of claim 1 , wherein the plurality of classifier networks may be 3D convolutional neural networks (CNN).

8. The method of claim 1 , wherein aligning the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks further comprises minimizing the distance between the correlation matrices during a training stage.

9. The method of claim 8 , wherein minimizing the distance between the correlation matrices further comprises minimizing the spatiotemporal semantic alignment loss.

10. A physical article of manufacture including one or more non-transitory computer-readable storage media, encoding computer-executable instructions for executing on a computer system a computer process, the computer process comprising:

generating a plurality of video streams of a sequence of object motion using a plurality of modalities;

training a plurality of classifier networks, each of the classifier network trained with a different one of the plurality of video streams, wherein each of the plurality of different classifier networks includes multiple intermediate layers;

determining correlation matrices of related intermediate layers of each of the plurality of the different classifier networks;

aligning the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks, wherein aligning the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks further comprises determining a spatiotemporal semantic alignment (SSA) loss between the related intermediate layers and by minimizing a distance between the correlation matrices; and

improving one of the plurality of classifier network's representation of a specific frame in its feature maps using the correlation matrices of another of the plurality of classifier networks.

11. The physical article of manufacture of claim 10 , wherein determining a spatiotemporal semantic alignment (SSA) loss between the related intermediate layers further comprises:

generating an adaptive regularization parameter; and

regularizing the spatiotemporal semantic alignment loss using the adaptive regularization parameter.

12. The physical article of manufacture of claim 11 , wherein generating the adaptive regularization parameter further comprising generating difference between classification losses of the plurality of the different classifier networks and generating the adaptive regularization parameter using the difference between classification losses.

13. The physical article of manufacture of claim 12 , wherein generating the adaptive regularization parameter further comprises setting the adaptive regularization parameter to zero to avoid negative knowledge transfer.

14. The physical article of manufacture of claim 10 , wherein the plurality of modalities includes two or more of RGB video modality, depth video modality, infrared video modality, and heat video modality.

15. The physical article of manufacture of claim 10 , wherein the plurality of classifier networks may be 3D convolutional neural networks (CNN).

16. The physical article of manufacture of claim 10 , wherein aligning the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks further comprises minimizing the distance between the correlation matrices in a training stage.

17. The physical article of manufacture of claim 16 , wherein minimizing the distance between the correlation matrices further comprises minimizing the spatiotemporal semantic alignment loss.

18. A physical hardware system, comprising:

a plurality of video devices to generate a plurality of video streams, each stream being of a different modality; and

a multimodal neural network system configured to:

train a plurality of classifier networks, each of the classifier network trained with a different one of the plurality of video streams, wherein each of the plurality of different classifier networks includes multiple intermediate layers;

determine correlation matrices of related intermediate layers of each of the plurality of the different classifier networks;

align the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks by minimizing a distance between the correlation matrices; and

improve one of the plurality of classifier network's representation of a one or more frames in its feature maps using the correlation matrices of the related intermediate layers of another of the plurality of classifier networks.

19. The physical hardware system of claim 18 , wherein the multimodal neural network system is further configured to align the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks by determining a spatiotemporal semantic alignment (SSA) loss between the related intermediate layers.

20. The physical hardware system of claim 18 , wherein the multimodal neural network system is further configured to align the correlation matrices of the related intermediate layers of each of the plurality of the different classifier networks by minimizing the distance between the correlation matrices by minimizing a spatiotemporal semantic alignment loss.