search

biubug6 / Pytorch_Retinaface

Retinaface get 80.99% in widerface hard val using mobilenet0.25.
MIT License
2.63k stars 774 forks source link

Having issue on reproducing results on WiderFace "hard" #210

Closed yakhyo closed 2 weeks ago

yakhyo commented 2 weeks ago

I refactored this repo with different feature extractor backbones. Refactored model weights perform well on WiderFace "easy" and "medium" but much lower on "hard" samples.

Here is my result when I use MXNet settings with mobilenetv1_0.25 (which is slightly better than current repo results)

Easy   Val AP: 0.8978211927262449
Medium Val AP: 0.8843002149567187
Hard   Val AP: 0.8228291929972594

This repo results:

Easy: 88.67% 
Medium: 87.09% 
Hard: 80.99%

However, when I use original image size for widerface evaluation my result for "hard" is worse than this repo:

Easy   Val AP: 0.9271427367537788
Medium Val AP: 0.9063996163204728
Hard   Val AP: 0.6224058241880429

This repo results:

Easy: 90.70% 
Medium: 88.16%
Hard: 73.82%

any suggestions or comments would be helpful. here is my re-implementation as a reference: https://github.com/yakhyo/retinaface-pytorch

yakhyo commented 2 weeks ago

I found the issue and reproduced the results.