New facial recognition dataset and optimization method announced by iQIYI
Chinese online entertainment service iQIYI has launched a new dataset and is presenting a white paper on enhancing the training of biometric facial recognition models with unlabeled data at the 2019 International Conference on Computer Vision (ICCV).
The paper “Unknown Identity Rejection Loss: Utilizing Unlabeled Data for Face Recognition” has been accepted by the ICCV, and puts forward a method of creating an optimization model for unlabeled data, which the iQIYI technical team says could help rapidly expand a facial recognition training database to improve the technology’s accuracy in uncontrolled settings.
At this year’s ICCV, iQIYI partnered with Imperial College London, DeepGlint and InfiniVision to hold a lightweight face recognition competition, during which iQIYI released its new iQIYI-VID-FACE dataset, which is made up of more than 10,000 celebrities in 6.3 million photos and hundreds of thousands of video clips from films and TV or frame images.
“A subject of increasing importance in AI research, multimodal person identification technology is a top priority for iQIYI researchers,” said Liu Junhui, iQIYI’s Senior Director. “Going forward, iQIYI will continue exploring the endless possibilities of entertainment experience by breaking through technological boundaries.”
The 2019 Celebrity Video Identification Challenge was jointly held by iQIYI and ACM International Conference on Multimedia, and was participated in by top universities and companies including Baidu, ZTE, and Nvidia.
ICCV 2019 is being held from October 27 to November 2 in Seoul, South Korea.
The origins of datasets used in facial recognition system training have been under increased scrutiny recently as public concerns about the privacy of online images grows.