ASFD: Automatic and Scalable Face Detector

• DOI: 10.48550/arxiv.2201.10781
• Publication Date: 2022-01-01
• Authors: Bin Zhang; Jian Li; Yabiao Wang; Ying Tai; Chengjie Wang; Jilin Li; Feiyue Huang; Yili Xia; Wenjiang Pei; Rongrong Ji
• Keywords: Margin (machine learning); Feature (linguistics); Video Graphics Array

Along with current multi-scale based detectors, Feature Aggregation and Enhancement (FAE) modules have shown superior performance gains for cutting-edge object detection. However, these hand-crafted FAE show inconsistent improvements on face detection, which is mainly due to the significant distribution difference between its training applying corpus, COCO vs. WIDER Face. To tackle this problem, we essentially analyse effect of data distribution, consequently propose search an effective architecture, termed AutoFAE by a differentiable architecture search, outperforms all existing in detection considerable margin. Upon found backbones, supernet further built trained, automatically obtains family detectors under different complexity constraints. Extensive experiments conducted popular benchmarks, Face FDDB, demonstrate state-of-the-art performance-efficiency trade-off proposed automatic scalable detector (ASFD) family. In particular, our strong ASFD-D6 best competitor AP 96.7/96.2/92.1 test, lightweight ASFD-D0 costs about 3.1 ms, more than 320 FPS, V100 GPU VGA-resolution images.