Multi-hop transformer for spatio-temporal reasoning and localization

1. A method for using a multi-hop reasoning framework to perform multi-step compositional long-term reasoning, the method comprising:

extracting feature maps and frame-level representations from a video stream by using a convolutional neural network (CNN);

performing object representation learning and detection;

linking objects through time via tracking to generate object tracks and image feature tracks;

feeding the object tracks and the image feature tracks to a multi-hop transformer that hops over frames in the video stream while concurrently attending to one or more of the objects in the video stream until the multi-hop transformer arrives at a correct answer; and

employing video representation learning and recognition from the objects and image context to locate a target object within the video stream.

2. The method of claim 1 , wherein spatial dimensions are collapsed into one dimension when performing the object representation learning and detection.

3. The method of claim 2 , wherein a batch dimension is combined with a temporal dimension for the feature maps when performing the object representation learning and detection.

4. The method of claim 3 , wherein positional encodings are learned for each time step and each resolution, and the feature maps are added in an element-wise manner to generate memory representations.

5. The method of claim 4 , wherein the memory representations are fed to a transformer-based object detector outputting transformed object representations used as inputs to a multilayer perceptron (MLP).

6. The method of claim 5 , wherein the transformer-based object detector accepts embeddings of object queries for every frame in the video stream.

7. The method of claim 6 , wherein tracking involves determining an association between two objects from two consecutive frames of the video stream defined by object class agreement and a difference between two bounding boxes.

8. The method of claim 1 , wherein multiple iterations over a single layer of the multi-hop transformer are run with a while loop.

9. The method of claim 8 , wherein every iteration of the multiple iterations performs one hop of reasoning by selecting target objects of the one or more objects in select frames of the video stream.

10. The method of claim 9 , wherein the hops operate in an auto-regressive manner.

11. A non-transitory computer-readable storage medium comprising a computer-readable program for using a multi-hop reasoning framework to perform multi-step compositional long-term reasoning, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:

extracting feature maps and frame-level representations from a video stream by using a convolutional neural network (CNN);

performing object representation learning and detection;

linking objects through time via tracking to generate object tracks and image feature tracks;

feeding the object tracks and the image feature tracks to a multi-hop transformer that hops over frames in the video stream while concurrently attending to one or more of the objects in the video stream until the multi-hop transformer arrives at a correct answer; and

employing video representation learning and recognition from the objects and image context to locate a target object within the video stream.

12. The non-transitory computer-readable storage medium of claim 11 , wherein spatial dimensions are collapsed into one dimension when performing the object representation learning and detection.

13. The non-transitory computer-readable storage medium of claim 12 , wherein a batch dimension is combined with a temporal dimension for the feature maps when performing the object representation learning and detection.

14. The non-transitory computer-readable storage medium of claim 13 , wherein positional encodings are learned for each time step and each resolution, and the feature maps are added in an element-wise manner to generate memory representations.

15. The non-transitory computer-readable storage medium of claim 14 , wherein the memory representations are fed to a transformer-based object detector outputting transformed object representations used as inputs to a multilayer perceptron (MLP).

16. The non-transitory computer-readable storage medium of claim 15 , wherein the transformer-based object detector accepts embeddings of object queries for every frame in the video stream.

17. The non-transitory computer-readable storage medium of claim 16 , wherein tracking involves determining an association between two objects from two consecutive frames of the video stream defined by object class agreement and a difference between two bounding boxes.

18. The non-transitory computer-readable storage medium of claim 11 , wherein multiple iterations over a single layer of the multi-hop transformer are run with a while loop.

19. The non-transitory computer-readable storage medium of claim 18 , wherein every iteration of the multiple iterations performs one hop of reasoning by selecting target objects of the one or more objects in select frames of the video stream.

20. A system for using a multi-hop reasoning framework to perform multi-step compositional long-term reasoning, the system comprising:

a memory; and

one or more processors in communication with the memory configured to:

extract feature maps and frame-level representations from a video stream by using a convolutional neural network (CNN);

perform object representation learning and detection;

link objects through time via tracking to generate object tracks and image feature tracks;

feed the object tracks and the image feature tracks to a multi-hop transformer that hops over frames in the video stream while concurrently attending to one or more of the objects in the video stream until the multi-hop transformer arrives at a correct answer; and

employ video representation learning and recognition from the objects and image context to locate a target object within the video stream.