search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
420 stars 77 forks source link

loss #22

Closed DD-BONED closed 2 years ago

DD-BONED commented 2 years ago

你好你使用了giouloss但是在这个任务中，c和(AuB)不应该是相等的吗

tzzcl commented 2 years ago

你说的对，实质上GIoU Loss在1D，并且中心点为同一个的时候，会退化成IoU Loss。

happyharrycn commented 2 years ago

Let me know if this question is now addressed.

DD-BONED commented 2 years ago

但是经过我打印发现len_c - unionk有些是非0的这个怎么理解

DD-BONED commented 2 years ago

但是经过我打印发现len_c - unionk有些是非0的这个怎么理解

happyharrycn commented 2 years ago

If you print the actual values of len_c - unionk, they are either zero or very close to zero (due to limited floating point precision).

DD-BONED commented 2 years ago

如果咋样，为什么还是要坚持是giou呢？

happyharrycn commented 2 years ago

With our current implementation, GIoU is indeed equivalent to IoU. One can either call our regression loss as GIoU or IoU. This equivalence, however, can not be generalized to the temporal action localization task, and is only due to the design of our model outputs. The GIoU loss has been proven to be more effective than IoU loss for regressing action boundaries. Should our design evolves, GIoU is preferred over IoU.

happyharrycn commented 2 years ago

Mark as resolved.