代码中关于Target refinement loss的计算和论文中不一致

MaverickRen / PixelLM

PixelLM is an effective and efficient LMM for pixel-level reasoning and understanding. PixelLM is accepted by CVPR 2024.

Apache License 2.0

147 stars 4 forks source link

代码中关于Target refinement loss的计算和论文中不一致 #10

Open jiangtann opened 4 months ago

jiangtann commented 4 months ago

文章中提到，对一个question中所有target的mask prediction probabilities求和后，如果某个pixel的probs >= 2，会增大其weights。即对那些预测出了多个targets的region增加惩罚权重。

但是代码中只对bg区域的权重置为0，并没有对预测出了多个targets的region增加权重： https://github.com/MaverickRen/PixelLM/blob/main/model/PixelLM.py#L95

虽然在93行得到了overlap_area，但是好像没有应用到计算loss的过程中。

jdg900 commented 4 months ago

Same here

MaverickRen commented 3 months ago

Sorry, this version of the code is incorrect. The non-overlapping area weight should be reset to 0 in the function 'overlap_loss'. I have modified this problem.