When taking videos as input

[hshustc]

When taking videos input, the feature maps in each layer have four dimensions, i.e., THW*C. Are the attention maps are still query-independent? Could you please give more details? Thanks a lot.

[tea1528]

Up-vote for this question. I am really interested whether the attention maps in video task showed similar result like object detection task.

Kinda think that temporal dimension should have some more importance over spatial dimensions.

[xvjiarui]

Sorry for the late reply

The attention across time is relative hard to visualize.

From the Table 1 in the paper, the attention on Kinetics seems to be a little more query dependent than COCO.
We will leave it as a future work.

[JJBOY]

In my experiments in video classification，non local moudle is not query-independent. How about you guys' results?