search

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

Some questions about small object detection #213

Closed weihancug closed 4 years ago

weihancug commented 4 years ago

I want to detect some vehicle objects from aerial images. The width of this kind of objects is about 10 ~ 20 pixels. My problems are as follows: 1) Do I need to apply the C2 layer of FPN to FCOS for small object detections? 2) If using C2 of FPN, do I only need to modify the FPN stride to [4, 8, 16, 32, 64,128] ?

Thanks for your kind assistance for solving my problems.

tianzhi0549 commented 4 years ago
  1. You can try to use C2.
  2. Yes, you need to change it.
weihancug commented 4 years ago

Thanks for your quickly replay. I have tried to increase the operation of stride 4, but found that the model is difficult to converge. The loss of the class and regression reduced to 0.2~0.3 and did not change. I just increase the number of convolutional body to 8 , which has a certain degree of effect. But the convergence is still slow.

Do you have any suggestions to speed up convergence

tianzhi0549 commented 4 years ago

@weihancug Please double check your code and make sure it is correct.

weihancug commented 4 years ago

Thanks for your help, I would check it.