发现rtmdet的一个bug

open-mmlab / mmdetection

OpenMMLab Detection Toolbox and Benchmark

https://mmdetection.readthedocs.io

Apache License 2.0

29.11k stars 9.38k forks source link

发现rtmdet的一个bug #9835

Open treasureqqm opened 1 year ago

treasureqqm commented 1 year ago

rtmdet_ins_head.py里面https://github.com/open-mmlab/mmdetection/blob/3.x/mmdet/models/dense_heads/rtmdet_ins_head.py#L920有一行代码错了，应该是self.reg_convs.append(reg_convs)而不是self.reg_convs.append(cls_convs)。论文里的模型是用这个代码训练的吗？这一行的错误预计会带来什么影响啊？

mm-assistant[bot] commented 1 year ago

We recommend using English or English & Chinese for issues so that we could have broader discussion.

RangiLyu commented 1 year ago

Thanks for reporting this bug. We will evaluate the performance impact of this bug as soon as possible.

RangiLyu commented 1 year ago

I trained RTMDet-Ins-s with separated cls reg heads. The box AP is 44.4(+0.3), however, the mask AP drops to 38.4(-0.3). It seems that fixing this bug has advantages and disadvantages for detection and instance segmentation. I'll further study the influence of the head structure and hope to maintain the mask AP.