Project for ECCV 2020 ChaLearn Looking at People Fair Face Recognition challenge

change based insightface

Requirements

1. python 2.7.12+

   • opencv
   • sklearn
   • numpy
   • scipy

2. mxnet >= 1.3.0

provided docker: zhaixingzhaiyue/mxnetcu90-py2

Instructions to reproduce the result of test phase with trained model

1. preprocess the test data

   Download the preprocessed data from the url https://c-t.work/s/0cc50e790e4f4e, and put it in the datasets folder. Then run tar xvf eccv_test_preprocessed.tar

   If you want preprocess the test data yourself:

   Download retina face model(https://pan.baidu.com/s/1C6nKq122gJxRhb37vK0_LQ) to facealign/RetinaFace/model

   cd facealign; sh ../tools/preprocess.sh # you need to change the eccv_test_data to real path in ../tools/preprocess.sh. Note, this is a little slow.

2. prepare the trained model

   Downlaod the trained model from https://c-t.work/s/9afa03a0bbb74c, and put it it the folder trained_models. Then run tar xvf trained_models.tar

3. generate final predict file

   sh ./tools/generate_sims.sh

   The final result will generate in the final_predictions directory with the name 'predictions.csv'

Instructions to train

1. train the model with ms1m dataset

   cd recognition/
   sh ./scripts/train_ms1m.sh

   stop at 185000 iterations

2. train the model with aligned ijbc

   cd recognition/
   sh ./scripts/train_ijbc_aligned.sh

   stop at 40000 iterations

3. train the model with origin ijbc

   cd recognition/
   sh ./scripts/train_ijbc_ori.sh

   stop at 40000 iterations