Lightweight Face Recognition Challenge

Abstract

Face representation using Deep Convolutional Neural Network (DCNN) embedding is the method of choice for face recognition. Current state-of-the-art face recognition systems can achieve high accuracy on existing in-the-wild datasets. However, most of these datasets employ quite limited comparisons during the evaluation, which does not simulate a real-world scenario, where extensive comparisons are encountered by a face recognition system. To this end, we propose two large-scale datasets (DeepGlint-Image with 1.8M images and IQIYI-Video with 0.2M videos) and define an extensive comparison metric (trillion-level pairs on the DeepGlint-Image dataset and billion-level pairs on the IQIYI-Video dataset) for an unbiased evaluation of deep face recognition models. To ensure fair comparison during the competition, we define light-model track and large-model track, respectively. Each track has strict constraints on computational complexity and model size. To the best of our knowledge, this is the most comprehensive and unbiased benchmarks for deep face recognition. To facilitate future research, the proposed datasets are released and the online test server is accessible as part of the Lightweight Face Recognition Challenge at the International Conference on Computer Vision, 2019.