Skip to content

VectXmy/FCOS.Pytorch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

FCOS: Fully Convolutional One-Stage Object Detection

implemented by pytorch1.0

Updates

  • ctr. on reg
  • giou loss
  • ctr. sampling

TODO

  • normalizing the regression targets

Requirements

  • opencv-python
  • pytorch >= 1.2
  • torchvision >= 0.4.

Anchor Points

Let's say the white boxes are the gt boxes, the points of different colors represent the sampling points of different feature layers while applying ctr-sampling.

Results

Train on 2 1080Ti, 3 imgs for each gpu, init lr=1e-5 cosine decays to 1e-6, but performance is not good on VOC07test. Maybe should remove centerness head while applying central sampling.
test1
test2
test3

Pretrained weights

Due to computational resource constraints, I was unable to fully train the model on the COCO dataset. I have converted the official pre-training model weights FCOS_R_50_FPN_1x into my own.
The converted weights is avaliable Baidu driver link, password: rpni
The official implementation of preprocessing(pixel is not normalized to 0-1 and input img follows BGR fomat ) is a little different from mine.

Other

some excellent work based on this repo:
FCOS-Pytorch-37.2AP
FCOS_DET_MASK

Releases

No releases published

Packages

No packages published

Languages