top of page
  • Jane Shen

ICCV 2019 | Pensees Won Three Championships of ICCV LFR

From October 27 to November 2, the biennial International Conference on Computer Vision (ICCV 2019) was held in Seoul, South Korea. ICCV, Conference on Computer Vision and Pattern Recognition (CVPR) and European Conference on Computer Vision (ECCV) are the three top conferences in the field of computer vision. Pensees’s first participation in the conference won itself three championships of ICCV LFR (Lightweight Face Recognition Challenge), and the Special Innovation Award issued by the organizing committee, which demonstrates Pensees’s application landing strength of computer vision technologies.

Unconstrained static image-based face recognition and dynamic video-based face recognition are one of the most in-depth research topics in the field of computer vision, and have a wide range of application scenarios in video surveillance, biomedical care and other fields. In recent years, new technologies and methods in various fields of computer vision have been proposed at many top conferences on computer vision, and face recognition technology based on deep learning has also made staged progress.

Although there are many research accomplishments in the field of face recognition, researches on lightweight face recognition based on deep learning still need to be strengthened. In light of face recognition applications with large-scale database, it is still a challenge to find a model of lightweight and high-precision that has excellent performance under unconstrained dynamic scenes of video surveillance.

ICCV LFR was set up for this purpose. It also became an important ICCV competition this year, attracting 292 teams from all over the world. Unlike other face recognition competitions, ICCV LFR posed strict limits on training data and test data. With such strict restrictions, to stand out from the 292 teams became extremely difficult.

ICCV LFR consisted of four competitions, each of which had its own restrictions and priorities:

Protocol-1 (DeepGlint-Light) lightweight image-based face recognition model, with computational complexity less than 1Gflops, model size less than 20MB, data type float32, and feature dimensionality 512 (FPR @ 1e-8);

Protocol-2 (DeepGlint-Large) large-scale image-based face recognition model, with computational complexity less than 30Gflops, data type float32, and feature dimensionality 512 (FPR @ 1e-8);

Protocol-3 (IQIYI-Light) lightweight video-based face recognition model, with computational complexity less than 1Gflops, data type float32, and feature dimensionality 512 (FPR @ 1e-4);

Protocol-4 (IQIYI-Large) large-scale video-based face recognition model, with computational complexity less than 30Gflops, data type float32, and feature dimensionality 512 (FPR @ 1e-4).

▲Oral presentation by Pensees Singapore Institute

In the end, the team led by Jane Shen Shengmei from Pensees Singapore Institute proposed an unsupervised learning method based on relational graphs to strengthen features, by which the team won the first place in the three competitions of ICCV LFR: lightweight image-based face recognition, large-scale image-based face recognition and lightweight video-based face recognition, far ahead of other research institutes and enterprises participating, including Microsoft Research Asia, Institute of Automation, Chinese Academy of Sciences,, and Toutiao. In the lightweight image-based face recognition competition, with a false positive rate (FPR) of one hundredth of a billion (1e-8), Pensees achieved a score of 93.41%, nearly six percentage higher than those of other participants. In the lightweight video-based face competition, Pensees’s 72.23% score was nearly nine percentage higher than those of other participants.

▲Results of the three ICCV LFR competitions

In the workshop of ICCV LFR, Pensees Singapore Institute gave an oral presentation entitled "A Graph Based Unsupervised Feature Clustering for Face Recognition", explaining the unsupervised learning method proposed by the team: relations of pair test data are used to make the feature distribution of data that belong to the same ID more compact, and feature distribution of data that belong to different IDs more disperse, thereby greatly improving recognition accuracy. The effectiveness of this method has been verified on IJB-C, YTF and CFP databases. In addition, the accuracy of the baseline model has been greatly improved.

▲Flow chart of unsupervised learning method based on relational graphs

Apart from winning three competitions of ICCV LFR, Pensees also broke the world record on IJB-C, the US NIST public face recognition dataset, with its latest algorithm model. IJB-C is the most scientific and comprehensive benchmark database in US NIST public face recognition datasets under unconstrained conditions. Considering the general accuracy saturation of LFW, CFP-FP, and AgeDB-30, IJB-C is currently a face recognition benchmark database containing video surveillance data which are most close to real-life scenarios.

▲Pensees’s test results on NIST IJB-C

In addition to the field of face recognition, Pensees has recently made breakthroughs in computer vision technologies such as pedestrian re-identification (ReID), video-based pedestrian re-identification (Video-based ReID), simultaneous localization and mapping (SLAM), winning world championships.

In July 2019, Pensees set a new world record on tests of three mainstream datasets of ReID: Market1501, DukeMTMC-reID, and CUHK03.

In August 2019, Pensees created the best results on tests of the three major datasets of Video-based ReID: PRID-2011, iLIDS-VID, and MARS.

In October, at ISMAR 2019, Pensees won the third place in the VSLAM competition of AR-SLAM Challenge. SLAM can be used in a variety of applications including autonomous driving, motion robot, 3D reconstruction, augmented reality and mixed reality. As an important accomplishment of Pensees in computer vision technologies, VSLAM shall add more value to the company's existing and future business. At the 17th China Public Security Exbo (CPSE) in Shenzhen, Pensees introduced an patrol robot for security scenarios. Later on, Pensees’s inhouse-developed and deep learning-based VSLAM technology will gradually replace 3D LiDAR for positioning and navigation of unmanned patrol cars, so as to reduce costs and expand application scenarios.

At present, Pensees has established research institutes in both Beijing and Singapore to continuously attract global AI talents. It has fully inhouse-developed, full-stack computer vision technologies, spanning multiple research fields of computer vision. On this basis, Pensees conducts scenario-oriented researches and innovations of AI technologies based on the company's existing business model and development direction. Furthermore, it continues its exploration of cutting-edge technologies from a global perspective, allowing the company to maintain sensitivity and attention to groundbreaking technologies, and to make good technical reserves for AI development and business exploration.

<The End>


77 views0 comments


bottom of page