question about offset gt generation

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)

https://arxiv.org/abs/1904.01355

Other

3.28k stars 630 forks source link

question about offset gt generation #154

Closed kfxw closed 5 years ago

kfxw commented 5 years ago

Hi! As shown in the following code (line 176-180) https://github.com/tianzhi0549/FCOS/blob/18191872ec9dff2eebd1560fe9e702a9f9d67714/fcos_core/modeling/rpn/fcos/loss.py#L176 I would like to know whether this operation will cause an entire box area splitted into different stages? And whether is that reasonable?

kfxw commented 5 years ago

To my understanding, such an operation means 'assign each pixel to a stage according to its max offset' rather than based on the longest bbox side as claimed in the paper.

tianzhi0549 commented 5 years ago

@kfxw yes. The same object might be assigned to different FPN levels. However, that is by design and aligned with the statement in our paper. Note that we hope that for each pixel, the whole object is included in its receptive field, so we choose to use the max offset.

kfxw commented 5 years ago

@tianzhi0549 Got it! Thx!