search

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
https://arxiv.org/abs/1904.01355
Other
3.27k stars 629 forks source link

hi,how to fix the problem that one pix belong to more than one class and these class belong to the same feature map? #306

Open gybuaa opened 4 years ago

gybuaa commented 4 years ago

Thanks for yours great job! According to the paper,we will just delete the sample with maximal area.And, the multi-level prediction can largely alleviate the problem.But how to fix the problem if one pix belong to more than one class and these classes belong to the same level feature map,and their area are the same？

tianzhi0549 commented 4 years ago

@gybuaa I think very few boxes are in this case. Just choose a random one in the case, which would not affect the final performance.

gybuaa commented 4 years ago

@gybuaa I think very few boxes are in this case. Just choose a random one in the case, which would not affect the final performance. Thanks a lot ！ may i ask if we want to avoid this problem absolutely,do you have any idea or research direction for me? thanks again for your reply！