Temporal multi-scale clockwork memory networks for object detection in videos
A computer-implemented method, system, and computer program product are provided for object detection utilizing an online flow guided memory network. The method includes receiving a plurality of videos, each of the plurality of videos including a plurality of frames. The method also includes generating, with a feature extraction network, a frame feature map for a current frame of the plurality of frames. The method additionally includes aggregating a memory feature map from the frame feature map and previous memory feature maps from previous frames on a plurality of time axes, with the plurality of time axes including a first time axis at a first frame increment and a second time axis at a second frame increment. The method further includes predicting, with a task network, an object from the memory feature map. The method also includes controlling an operation of a processor-based machine to react in accordance with the object.
1. A computer-implemented method for object detection utilizing a clockwork flow guided memory network, the method comprising:
receiving, by a processor, a plurality of videos, each of the plurality of videos including a plurality of frames;
generating, by the processor with a feature extraction network, a frame feature map for a current frame of the plurality of frames;
aggregating, by the processor, a memory feature map determined from the frame feature map and previous memory feature maps from previous frames on a plurality of time axes by warping the previous memory feature maps, with the plurality of time axes including a first time axis at a first frame increment and a second time axis at a second frame increment with the second frame increment greater than the first frame increment, the warping being updated at each of the plurality of frames and being configured to perform a single-warp computation per-frame;
predicting, by the processor with a task network, an object from the memory feature map; and
controlling an operation of a processor-based machine to react in accordance with the object.
2. The computer-implemented method as recited in claim 1 , wherein the aggregating includes utilizing a third time axis at a third frame increment in the plurality of time axes, with the third frame increment greater than the second frame increment.
3. The computer-implemented method as recited in claim 1 , wherein the aggregating includes utilizing a displacement network.
4. The computer-implemented method as recited in claim 3 , wherein the displacement network includes an optical flow convolutional neural network.
5. The computer-implemented method as recited in claim 3 , wherein the displacement network generates displacement vectors from the current frame and the previous frame.
6. The computer-implemented method as recited in claim 1 , wherein the aggregating includes combining the frame feature map with a warped memory feature map.
7. The computer-implemented method as recited in claim 6 , wherein the combining includes utilizing simple averaging.
8. The computer-implemented method as recited in claim 6 , wherein the combining includes employing learned adaptive weighting.
9. The computer-implemented method as recited in claim 6 , wherein the aggregating further includes warping one of the previous memory feature maps with displacement vectors from a displacement network to form the warped memory feature map.
10. The computer-implemented method as recited in claim 9 , wherein the warping includes employing bilinear sampling.
11. The computer-implemented method as recited in claim 1 , wherein the task network includes a task convolutional neural network.
12. The computer-implemented method as recited in claim 1 , wherein one of the plurality of time axes utilizes the plurality of frames in a continuous order.
13. The computer-implemented method as recited in claim 1 , wherein one of the plurality of time axes utilizes the plurality of frames in an order that skips frames.
14. The computer-implemented method as recited in claim 1 , wherein the aggregating includes utilizing a plurality of aggregation systems.
15. The computer-implemented method as recited in claim 14 , wherein one of the plurality of aggregation systems combines the frame feature map with the previous memory feature map from the previous frame from each of the plurality of time axes.
16. The computer-implemented method as recited in claim 15 , wherein the previous frame from one of the plurality of time axes is more than one frame away from the current frame.
17. The computer-implemented method as recited in claim 1 , wherein the aggregating includes utilizing an aggregation system for each of the plurality of time axes.
18. The computer-implemented method as recited in claim 1 , wherein the plurality of time axes includes more than three time axes.
19. A computer program product for object detection utilizing a clockwork flow guided memory network, 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:
receiving, by a processor, a plurality of videos, each of the plurality of videos including a plurality of frames;
generating, by the processor with a feature extraction network, a frame feature map for a current frame of the plurality of frames;
aggregating, by the processor, a memory feature map determined from the frame feature map and previous memory feature maps from previous frames on a plurality of time axes by warping the previous memory feature maps, with the plurality of time axes including a first time axis at a first frame increment and a second time axis at a second frame increment with the second frame increment greater than the first frame increment, the warping being updated at each of the plurality of frames and being configured to perform a single-warp computation per-frame;
predicting, by the processor with a task network, an object from the memory feature map; and
controlling an operation of a processor-based machine to react in accordance with the object.
20. An object detector that detects objects in video, the object detector comprising:
a processing system including a processor device and memory coupled to the processor device, the processing system having program code configured to:
receive a plurality of videos, each of the plurality of videos including a plurality of frames;
generate, with a feature extraction network, a frame feature map for a current frame of the plurality of frames;
aggregate a memory feature map determined from the frame feature map and previous memory feature maps from previous frames on a plurality of time axes by warping the previous memory feature maps, with the plurality of time axes including a first time axis at a first frame increment and a second time axis at a second frame increment with the second frame increment greater than the first frame increment, the warping being updated at each of the plurality of frames and being configured to perform a single-warp computation per-frame;
predict, with a task network, an object from the memory feature map; and
control an operation of a processor-based machine to react in accordance with the object.