研究与发表
RetinaFace
Single-stage Dense Face Localisation in the Wild
RetinaFace combines accurate face detection with reliable five-point landmark localisation, making it a practical front-end for recognition and alignment-heavy pipelines.
论文信息
RetinaFace: Single-stage Dense Face Localisation in the Wild
发表渠道
CVPR 2020
作者
Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, Stefanos Zafeiriou
研究概览
RetinaFace focuses on detecting difficult faces in real scenes while also returning landmarks that improve alignment quality. That combination makes it useful not only for detection benchmarks, but also for enterprise systems where downstream recognition accuracy depends on consistent cropping and pose normalization.
落地应用
- Face detection and alignment before recognition
- Real-time video analytics and entry gate cameras
- Image ingestion, cropping, and portrait normalization
- Mobile capture flows that require stable landmark estimation
核心贡献
Jointly predicts face boxes and five facial landmarks in a single-stage detector, reducing pipeline complexity.
Dense supervision and context-aware design improve robustness on small, occluded, blurred, and profile faces.
Provides dependable alignment signals for recognition, quality assessment, and video analytics workflows.