Custom Centernet for face detection and other object detection tasks.

In this repo, you can training object detection tasks with many other backbones not only the backbones in this repo. Refered linknet(For segmentation), I desigined the linknet backend model for object detection. Since I am used to use the voc format dataset(the data is always marked by LabelImg), the voc format dataset has been provided. Anyway, you could also use the centernet original coco format to train the model. For application, I have provided the python shell to convert the *.pth to onnx.

Requirements for this repo

• python>=3.6
• Pytorch>=1.1
• cocoapi
• opencv-python
• Cython
• numba
• progress
• matplotlib
• easydict
• scipy

Anaconda 3 is recommended

Training

Wilderface Data Prepare

Refered to the Ultra-Light-Fast-Generic-Face-Detector-1MB , Generate the voc format wilderface dataset and copy it to the ./data And you can use it to train or generate the json format annotations for coco format

Training wilderface with centerface's structure(relu) without keypoints

in the ./src/main.py : from opts_voc import opts or from opts_coco import opts cd ./src/, python main.py

Training wilderface with centerface's structure(relu) with keypoints

• (1) heatmap way just like the original CenterNet's person keypoints If you want to train the model with keypoints , you need to generate two json format annotations for coco format in the data/wilderface/ as follows:wildertrain.json and wilderval.json by the VOC2COCOtrain.py and VOC2COCOval.py in the ./data dir . in the ./src/main.py : from opts_hplm import opts cd ./src/, python main.py

• (2) Just directly regress keypoints by coordinates Actually,the Centerface's keypoints regression in the ./src/main.py : from opts_reglm import opts cd ./src/, python main.py

Inference demo

In the root_dir/src/demo.py , import relative opts(eg. from opts_voc import opts ) python demo.py --demo [img|video|webcam] --load_model [model_path,../models/model_best.pth] (just like the original centernet)

Wilderface AP (easy ,medium and hard)

root_dir/wilderface_eval_space Python Evaluation Code for Wider Face Dataset refered to Ultra-Light-Fast-Generic-Face-Detector-1MB

• run evaluation_on_widerface.py import relative opts(eg. from opts_voc import opts ) in the evaluation_on_widerface.py python evaluation_on_widerface.py

• before evaluating python setup.py build_ext --inplace

• evaluating GroungTruth: wider_face_val.mat, wider_easy_val.mat, wider_medium_val.mat,wider_hard_val.mat python evaluation.py -p -g

• the pretrained model's ap on the wilderface (face confidence>0.4 iouthreshold=0.4)

resolution	Easy Set	Medium Set	Hard Set
640x640	0.872	0.839	0.655
800x800	0.855	0.836	0.715

Convert model to onnx

import relative opts(eg. from opts_voc import opts ) in the evaluation_on_widerface.py in the opts_voc.py ,--load_model (your trained model) root_dir/src/pytorch2onnx.py python pytorch2onnx.py , and the onnx model is in the onnxmodel dir

For tensorRT's inference C++, you can refer to TensorRT-CenterNet

TODO

• Some Efficient pretrained models and mAP report in the wilderface
• Application demos by convert the onnx to other formats (mnn,ncnn,caffe for HUAWEI' nnie and npu)

Reference

• Most refered to CenterNet byxingyizhou

• Centerface byStar-Clouds

• Ultra-Light-Fast-Generic-Face-Detector-1MB byLinzaer

• CenterFace.pytorch by chenjun2hao

• CenterMulti by bleakie

• TensorRT-CenterNet by CaoWGG