Tsinghua University | webface260m: benchmark for million level deep face recognition (tpami2022)

Paper title ：WebFace260M: A Benchmark for Million-Scale Deep Face Recognition

Thesis link ：https://arxiv.org/abs/2204.10149

Author's unit ： Tsinghua University & Core technology & Imperial College of Science, Technology and Medicine

Face benchmark enables the research community to train and evaluate high-performance face recognition systems . In this paper , We provide a new million level identification benchmark , Contains unprocessed 4M identity /260M Face (WebFace260M) And clean 2M identity /42M Face (WebFace42M) Training data , And a well-designed time constraint evaluation protocol . First , We have collected 4M And from Internet Downloaded 260M Face of . then , An automatic self-training method is designed (CAST) Clean the pipeline to purify the huge WebFace260M, It is efficient and scalable . As far as we know , After cleaning WebFace42M It is the largest public face recognition training set , We hope to narrow the data gap between academia and Industry . Refer to the actual deployment , A face recognition algorithm with reasoning time constraints is constructed （FRUITS） Protocol and a new test set with rich attributes . Besides , We collected a large collection of masked faces , be used for COVID-19 Biometric assessment under . In order to comprehensively evaluate the face matcher , Respectively in the standard 、 Perform three recognition tasks under masked and unbiased settings . With the help of this benchmark , We have deeply studied the million level face recognition problem . A distributed framework is developed to effectively train face recognition models without tampering with performance . stay WebFace42M With the support of , We are in a challenging IJB-C Reduced on set 40% The failure rate of , And in NIST-FRVT Of 430 Ranked third among the items . Compared with the public training set , Even if it's 10% The data of (WebFace4M) It also shows excellent performance . Besides , stay FRUITS-100/500/1000 A comprehensive baseline is established under the millisecond protocol . The proposed benchmark is in the standard 、 Masked and unbiased face recognition scenarios show great potential .