Nanodet - Face Detecion

Nanodet trained models for Face Detection. For more details check out Nanodet repo. Models were trained with Google Colab servers (Single GPU)

Directory structure

.
├── face                                 # Face related object detection - OpenImages Dataset (20k train : 4k test)
│   ├── nanodet_m_1.0x_sgd                      # Cfg, checkpoints, ... for nanodet_m_1.0x @ 320x320
│   ├── nanodet_m_0.5x_sgd_416x416              # Cfg, checkpoints, ... for nanodet_m_0.5x @ 416x416
|   ├── nd-efficientnet_lite0_more_aug_320x320  # Cfg, checkpoints, ... for efficientnet_lite0 @ 320x320
└── ...

Note: Current models are trained with 20k images (Human faces) of Open Images v6. All these faces satisfy: bbox_area/image_area > 0.03. So expect poor performance on those images which bbox occupy less than 3% of the image.

TODO

• ncnn optimized models
• EfficientNet-Lite0 model
• Yolo Fastest v2