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	Head	Data	LFW	CFP-FP	CPLFW	AGEDB	CALFW	IJBB (TPR@FAR=1e-4)	IJBC (TPR@FAR=1e-4)
IR_101	ArcFace	MS1Mv2	99.77	98.27	92.08	98.15	95.45	94.2	95.6
IR_101	CurricularFace	MS1Mv2	99.80	98.36	93.13	98.37	96.05	94.86	96.15
IR_18	ArcFace	MS1Mv2	99.65	94.89	89.80	97.23	95.60	90.06	92.39
IR_34	ArcFace	MS1Mv2	99.80	97.27	91.75	98.07	95.97	92.88	94.65
IR_50	ArcFace	MS1Mv2	99.80	97.63	92.50	97.92	96.05	93.45	95.16
MobileFaceNet	ArcFace	MS1Mv2	99.52	91.66	87.93	95.82	95.12	87.07	89.13
GhostNet_x1.3	ArcFace	MS1Mv2	99.65	94.20	89.87	96.95	95.58	89.61	91.96
EfficientNetB0	ArcFace	MS1Mv2	99.60	95.90	91.07	97.58	95.82	91.79	93.67
EfficientNetB1	ArcFace	MS1Mv2	99.60	96.39	91.75	97.65	95.73	92.43	94.43

Backbone model size & latency

The device and platform information see below:

	Device	Inference Framework
x86 cpu	Intel(R) Xeon(R) Platinum 8255C CPU @ 2.50GHz	Openvino
arm	Kirin 980	TNN

Test results for different backbones and different devices：

Backbone	Model Size(fp32)	X86 CPU	ARM
EfficientNetB0	16MB	26.29ms	32.09ms
EfficientNetB1	26MB	35.73ms	46.5ms
MobileFaceNet	4.7MB	7.63ms	15.61ms
GhostNet_x1.3	16MB	25.70ms	27.58ms
IR_18	92MB	57.34ms	94.58ms
IR_34	131MB	105.58ms	NA
IR_50	167MB	165.95ms	NA
IR_101	249MB	215.47ms	NA

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