deepinx / iqiyi-vid-challenge

Code for IQIYI-VID Challenge based on ESSH and Insightface

Recently www.iqiyi.com released a great video person dataset called IQIYI_VID and also launched a person search competition on it. It is a very large and real dataset worth trying to verify your face model accuracy precisely.

This repository contains the code for IQIYI-VID(IQIYI video person identification) Challenge. The methods are implemented in Python and MXNet. The Enhanced SSH (ESSH) from enhanced-ssh-mxnet is applied for face detection and alignment. Insightface scheme is used for face recognition.

Pre-trained models can be downloaded on BaiduCloud or GoogleDrive.

Environment

This repository has been tested under the following environment:

• Python 2.7
• Ubuntu 18.04
• Mxnet-cu90 (==1.3.0)

Installation

1. Prepare the environment.

2. Clone the repository.

3. Type make to build necessary cxx libs.

4. Download the pre-trained model and place it in ./model/

5. Download the IQIYI-VID Datasets from IQIYI_VID and unzip them to data/iqiyi_vid directory.

Usage

1. Detect faces on train+val dataset and test dataset respectively using ESSH model. Model model-r50-gg is used to judge the quality of the detected faces.

python detect.py --model ./model/model-r50-gg/model,0 --output ./output/det_trainval --dataset ./data/iqiyi_vid --gpu 0 --stage trainval
python detect.py --model ./model/model-r50-gg/model,0 --output ./output/det_test --dataset ./data/iqiyi_vid --gpu 0 --stage test

2. Extract features to the detected faces of train+val and test dataset respectively using model-r100-gg model.

python feature.py --model ./model/model-r100-gg/model,0 --input ./output/det_trainval --output ./output/feat_trainval  --gpu 0
python feature.py --model ./model/model-r100-gg/model,0 --input ./output/det_test --output ./output/feat_test --gpu 0

3. Re-save the extracted face features for training the MLP network.

python genfeat.py --inputs ./output/feat_trainval --output ./output/trainval

4. Train the MLP network for face ID recognition using train+val datasets.

python train_mlp.py --data ./output/trainval --prefix ./model/iqiyi --ckpt 1 --network r50 --lr 0.2 --per-batch-size 1024

5. Predict face ID from features of the test dataset using the pre-trained MLP network.

python predict.py --model ./model/iqiyi,40 --gpu 0 --inputs ./output/feat_test --output ./output/pred_test

6. Run python submit.py to generate the final submissions for IQIYI-VID Challenge.

License

MIT LICENSE

Reference

@article{deng2018arcface,
title={ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
author={Deng, Jiankang and Guo, Jia and Niannan, Xue and Zafeiriou, Stefanos},
journal={arXiv:1801.07698},
year={2018}
}

@inproceedings{Najibi2017SSH,
  title={SSH: Single Stage Headless Face Detector},
  author={Najibi, Mahyar and Samangouei, Pouya and Chellappa, Rama and Davis, Larry S.},
  booktitle={IEEE International Conference on Computer Vision},
  year={2017},
}

Acknowledgment

The code is adapted based on an intial fork from the insightface repository.