A trusty face recognition research platform developed by Tencent Youtu Lab

Introduction

TFace: A trusty face recognition research platform developed by Tencent Youtu Lab. It provides a high-performance distributed training framework and releases our efficient methods implementation.

This framework consists of several modules: 1. various data augmentation methods, 2. backbone model zoo, 3. our proposed methods for face recognition and face quality, 4. test protocols of evalution results and model latency.

Recent News

2021.3: SDD-FIQA: Unsupervised Face Image Quality Assessment with Similarity Distribution Distance accepted by CVPR2021. [paper] [code]

2021.3: Consistent Instance False Positive Improves Fairness in Face Recognition accepted by CVPR2021. [paper] [code]

2021.3: Spherical Confidence Learning for Face Recognition accepted by CVPR2021. [paper] [code]

2020.8: Improving Face Recognition from Hard Samples via Distribution Distillation Loss accepted by ECCV2020. [paper] [code]

2020.3: Curricularface: adaptive curriculum learning loss for deep face recognition has been accepted by CVPR2020. [paper] [code]

Requirements

• python==3.6.0
• torch==1.6.0
• torchvision==0.7.0
• tensorboard==2.4.0
• Pillow==5.0.0

Getting Started

Train Data

The training dataset is organized in tfrecord format for efficiency. The raw data of all face images are saved in tfrecord files, and each dataset has a corresponding index file(each line includes tfrecord_name, trecord_index offset, label).

The IndexTFRDataset class will parse the index file to gather image data and label for training. This form of dataset is convenient for reorganization in data cleaning(do not reproduce tfrecord, just reproduce the index file).

1. Convert raw image to tfrecords, generate a new data dir including some tfrecord files and a index_map file

python3 tools/img2tfrecord.py --img_list=${img_list} --pts_list=${pts_list} --tfrecords_name=${tfr_data_name}

2. Convert old index file(each line includes image path, label) to new index file

python3 tools/convert_new_index.py --old=${old_index} --tfr_index=${tfr_index} --new=${new_index}

3. Decode the tfrecords to raw image

python3 tools/decode.py --tfrecords_dir=${tfr_dir} --output_dir=${output_dir}

Augmentation

Data Augmentation module implements some 2D-based methods to generated some hard samples, e.g., maks, glass, headscarf. Details see Augmentation

Train

Modified the DATA_ROOTandINDEX_ROOTin ./tasks/distfc/train_confing.yaml, DATA_ROOT is the parent dir for tfrecord dir, INDEX_ROOT is the parent dir for index file.

bash local_train.sh

Test

Detail codes and steps see Test

Benchmark

Evaluation Results

Backbone model size & latency

The device and platform information see below:

Test results for different backbones and different devices：

Acknowledgement

This repo is modified and adapted on these great repositories, we thank theses authors a lot for their greate efforts.

• cavaface.pytorch
• face.evoLVe.PyTorch
• insightface
• mobile-vision