RetinaFace Face Detector

Introduction

RetinaFace is a practical single-stage SOTA face detector which is initially introduced in arXiv technical report and then accepted by CVPR 2020.

This is a modified repository structured similarly to the original repository which can be found here. Group-F has tried to recreate the results obtained by the original code on the WiderFace dataset as well as obtain results on a new dataset created by us.

Original Data

1. Download the WiderFace dataset from the drive link here. Create a folder data and unzip the contents in that folder.
2. Download the pre-trained model from the drive link here. Create a folder model and unzip the contents in that folder.
3. Execute the make file to setup the environment.
4. Using a conda environment, install additional dependencies, pip install mxnet-cu101 opencv-python easydict Cython.
5. Use the command python test_widerface.py --prefix './model/R50' --mode=1 to test the pretrained RetinaFace-50 (R50) model on the WiderFace dataset. The outputs will be saved in the wout folder.
6. Move the wout folder in eval/eval_tools and execute the wider_eval.m file to produce the output PR curve graphs.

New Data

1. The new dataset is present in the new_data folder. It consists of 3 subsets along with a combined set of all images and label files.
2. To test RetinaFace-50 on this new data, run the command python test.py new_data/all/img output/all. This will produce the labels as predicted by RetinaFace-50 in the output folder.
3. Run python test_output.py output/all/text new_data/all/text all to produce the PR curves along with MAP values for the new data.