search

Epiphqny / SOLOv2

SOLOv2: Dynamic, Faster and Stronger, achives 39.5mAP on coco test-dev (36 epochs result)
https://arxiv.org/abs/2003.10152
Apache License 2.0
245 stars 37 forks source link

Question about mask loss #7

Closed qin2294096 closed 4 years ago

qin2294096 commented 4 years ago

Thanks for your work!

https://github.com/Epiphqny/SOLOv2/blob/master/mmdet/models/anchor_heads/solov2_head.py#L373 https://github.com/Epiphqny/SOLOv2/blob/master/mmdet/models/anchor_heads/solov2_head.py#L200 I have noticed that, in training, the gt masks and mask predictions are downsampled to smaller scales(1/4, 1/4, 1/8, 1/16, 1/16 from p2~p6). And these downsampled masks are used to compute loss.

However, solov2 can predict masks on 1/4 scale directly, so my question is, why not compute loss with gt masks on 1/4 scale like yolact? Thank you. （为什么这里不像yolact那样，直接用1/4大小的gt mask去计算损失？）

Epiphqny commented 4 years ago

@qin2294096 I have implemented the 1/4 scale mask loss before, the result is about 0.2 point higher than current version, because the code is similar, i didn't update.