This is my first project at Cyber Core Co.,LTd., leading by Mr. Khai Nguyen Quoc
The model is trained on 300W-LP dataset and test on AFLW2000 dataset.
New approach: crop only face by RetinaFace (ResNet50) and use it to train model (old method use raw images).
Our network beats the state-of-the-art FSANet on AFLW2000 dataset
The table bellow show our result. Speed is tested on GPU GEFORCE GTX 1080 Ti.
| Method | MB | Yaw | Pitch | Roll | MAE average | Inference Time (ms) |
|---|---|---|---|---|---|---|
| Dlib (68 points) | - | 23.1 | 13.6 | 10.5 | 15.73 | |
| FAN (12 points) | 183 | 6.36 | 12.3 | 8.71 | 9.12 | |
| Landmarks | - | 5.92 | 11.86 | 8.27 | 8.68 | |
| 3DDFA | - | 5.4 | 8.53 | 8.25 | 7.39 | |
| Hopenet (α=2) | 95.9 | 6.47 | 6.56 | 5.44 | 6.16 | |
| Hopenet (α=1) | 95.9 | 6.92 | 6.64 | 5.67 | 6.41 | |
| SSR-Net-MD | 1.1 | 5.14 | 7.09 | 5.89 | 6.04 | |
| FSA-Caps-Fusion | 5.1 | 4.5 | 6.08 | 4.64 | 5.07 | 1 (size 64x64) |
| Our method (ResNet18) | 49.00 | 4.13 | 5.70 | 4.33 | 4.72 | 3 (size 128x128) |
