A PyTorch implementation of RetinaFace: Single-stage Dense Face Localisation in the Wild. The official code in Mxnet can be found here.
[ ] Move to detectron2
| Style | easy | medium | hard |
|---|---|---|---|
| Mxnet(Single Scale 2150) | 88.72% | 86.97% | 79.19% |
| Mxnet(Original Scale) | 89.58% | 87.11% | 69.12% |
| Pytorch(Same as Mxnet 2150) | 91.795% | 90.49% | 85.55% |
| Pytorch(Original Scale) | 92.65% | 90.91% | 80.23% |
- pytorch == 1.3.1
- torchvision == 0.4.0
- python == 3.7
cd libs/DCNv2
./make.shpython3 demo.py