tkianai/retinaface.detectron2

Implement RetinaFace based on detectron2/使用detectron2框架实现retinaface

RetinaFace.detectron2

RetinaFace detector based on detectron2.

TODO

• Baseline R-50
• Baseline mnet[v1, v2]
• Update the Conv and Norm Used
• Update to suitable for latest detectron2 version: remove duplicate anchors
• Add DFConv
• Add BiFPN
• Add other backbones
• Add visual samples

Usage

Train a new model with this project

• Stylize the original widerface dataset to coco-style

python tools/widerface_stylized_coco.py 

• Train

python train_net.py --num-gpus <number of gpus> --config-file <your config file>

For example, train with a baseline with backbone(resnet50), you can use this codes:

python train_net.py --num-gpus 4 --config-file configs/retinaface/retinaface_R_50_FPN.yaml

Test a trained model with this project

When testing model, you only need to specify the following essential arguments:

• config-file: the same with the config file when you training the model
• input: video path or regular match file path
• output: specify the output folder to save the result
• opts: loading the latest trained checkpoint

For example,

python demo.py --config-file configs/retinaface/retinaface_R50_FPN.yaml --input datasets/widerface/val/images/*/*.jpg --output work_dirs/retinaface_R_50_FPN/val  --opts MODEL.WEIGHTS work_dirs/retinaface_R_50_FPN/model_final.pth

Evaluate a trained model using the widerface toolsets

After generated the result(Test part),

• Change the results format into widerface style

python tools/predictions_to_widerface_val.py --res work_dirs/retinaface_R_50_FPN/val/results.pkl --save work_dirs/retinaface_R_50_FPN/widerface

• Use widerface_evaluate to visual the performance

cd tools/widerface_evaluate
python evaluation.py --p ../../work_dirs/retinaface_R_50_FPN/widerface