Skip to content

Lightweight face detection models archieve high peformance on widerface val set

Notifications You must be signed in to change notification settings

HamsterLC/RetinaFace_enhance

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

52 Commits
 
 
 
 

Repository files navigation

This repo is an improvement version of https://github.com/biubug6/Pytorch_Retinaface

Improvements in is repo:

  1. More Data Augmentation(Expand, MinIOU Crop with different ratio, Multi-Scale Training)
  2. Long Training Schedule with more epoch and stage
  3. Support More Backbone(MV1_0.25, MV2_0.35, MV2_1.0)

Comparison with origin implement(original image scale)

Style BackBone MAdds Parameters easy medium hard
Enhance_MV1 MV1_0.25 1.0B 0.44M 91.8% 90.2% 76.4%
Enhance_MV2 MV2_0.35 1.0B 0.48M 92.8% 90.9% 79.4%
RetinaFace_Ori[6] MV1_0.25 1.0B 0.44M 89.6% 87.1% 69.1%
RetinaFace_biubug6[7] MV1_0.25 1.0B 0.44M 90.7% 88.2% 73.8%

AP on hard set can surpass 80% by using knowledege distilling. Some deploy models are in the Deploy folder.

Reference

  1. MobileNetV2: Inverted Residuals and Linear Bottlenecks
  2. Ultra-Light-Fast-Generic-Face-Detector-1MB
  3. libfacedetection
  4. ASFD: Automatic and Scalable Face Detector
  5. CenterFace
  6. RetinaFace(insightface)
  7. RetinaFace(biubug6)
  8. RetinaFace(lbin)

About

Lightweight face detection models archieve high peformance on widerface val set

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages