Blazeface trained on pascal_voc person.
This repo is an unofficial implementation of:
BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs
Paper
The repo contains..
- SSD-like model
- Training script
- Eval script
- Focal loss
- Trained weights
- CenterNet-like detection for more accuracy and faster opereations. (Under dev).
The training and inference is > 10x faster than SSD.
The BlazeFace model is based on here, and localaization and detection layers are added from ssd.pytorch.
Also the nms functions are borrowed from ssd.pytorch as well.
Thank you!
Pytorch > 1.0
opencv
scikit-learn
Run:
python train_BlazeFace.py
See train_BlazeFace.ipynb for specifics.
Dataset_test_with_BlazeFace128-VOC-allclasses.ipynb runs all VOC classes, but doesn't get good accuracy at this point.
See inference.ipynb to run inference.
It takes less than 30ms on K80 to run single image (including nms).