multi-class nms about actionformer_release HOT 5 CLOSED

Hi, I am not sure that I fully understand your questions.

But our code provides a simple choice to enable/disable multi-class nms, you may refer to L47-L50 in THUMOS config for instructions.

from actionformer_release.

Hi, I am not sure that I fully understand your questions.

But our code provides a simple choice to enable/disable multi-class nms, you may refer to L47-L50 in THUMOS config for instructions.

不好意思，可能英文没表达清楚，还是直接说中文吧。
mulit-class设置成false其实也会存在这样的情况：假设两个预测的结果分别是[ 1秒-3秒，类别为1] [1秒-1.1秒，类别为2]，这两个预测结果的iou比较小, nms的iou阈值也并不能过滤掉这种情况，那最终1秒-1.1秒这个区间，类别应该怎么定呢？其实主要想问的问题就是可能存在有一段区间，但是属于两个不同的label的预测结果，应该怎么处理比较好？

from actionformer_release.

Similar to object detection, temporal action localization is a structured prediction problem, where the output is a set of temporal actions (events) rather than frame-level labels. In your example here, the output does not have any ambiguity. The model predicts an action that starts at 1s and ends at 4s with category 1, and another action that starts at 2s and ends at 5s with category 2. The segment starting at 2s and ending at 5s is not considered as an action by the model. This has nothing to do with NMS.

from actionformer_release.

Let me know if there is any further questions. Otherwise, I will mark the issue as resolved.

from actionformer_release.

Resolved due to inactivity.

from actionformer_release.