Unexpected cls_targets results for short segments in label_points_single_video

[krkrkrrk]

Hello.

In the label_points_single_video method, when the integer parts of the segment's start and end values are the same (e.g., from 4.1 to 4.3) (this refers to segments, not timestamps), all elements of inside_gt_seg_mask become False, resulting in all elements of cls_targets being 0. Is this a bug? If this behavior is as intended, could you explain why it is implemented this way?

For example, in the case of gt_segment = (4.1, 4.3), I believe cls_targets[5, 0] should be 1. However, with the current implementation, all values become 0.

I also have a similar question. For instance, in the case of gt_segment = (3.3, 4.7), I think both cls_targets[4, 0] and cls_targets[5, 0] should be 1, but only cls_targets[5, 0] becomes 1. Am I misunderstanding something? Could you explain why it is implemented this way?

[tzzcl]

For the groundtruth assignment stage, you can refer to the center sampling paragraph in the paper. In short, short actions may be ignored and do not have groundtruth cls_target and reg_targets.

[krkrkrrk]

Thank you for your quick reply.

I’m sorry, but I don’t understand the relationship between "center sampling" and "short actions may be ignored".

From my understanding, center sampling selects the regions near the center of an action as positive and has no relation to ignoring short actions. So, I think my issue might not be related to center sampling. Could you please explain it in more details?

Also, If center sampling causes short actions to be ignored, what changes can be made to ensure that short actions are not ignored?

I look forward to your reply.