Improved Selective Refinement Network (ISRN): JD AI’s State‑of‑the‑Art Face Detection Achieves Top Rankings on WIDER FACE

JD AI Research Institute's Computer Vision and Multimedia Lab recently presented the Improved Selective Refinement Network (ISRN) for face detection, achieving first‑place results on the WIDER FACE benchmark’s Easy, Medium, and Hard tracks with AP scores of 96.3%, 95.4% and 90.3% respectively.

The WIDER FACE dataset, hosted by the Chinese University of Hong Kong, is one of the largest and most challenging public face‑detection collections, containing 32,203 images and 393,703 annotated faces with extreme variations in scale, pose, occlusion, expression, illumination, and makeup.

ISRN builds on the earlier Selective Refinement Network (SRN) by introducing several key improvements:

Random initialization of a large number of network architectures combined with Group Normalization, eliminating the need for traditional ImageNet pre‑training.

Pre‑training on the MS COCO dataset, which provides abundant human‑class images and better scale diversity for small‑face detection.

Redesign of the residual network input module to avoid down‑sampling, preserving fine‑grained facial location information, and splitting the first multi‑channel convolution to keep computational cost low.

Enhanced small‑face detection via Feature Pyramid Network (FPN) fusion, together with Selective Two‑step Classification (STC) and Selective Two‑step Regression (STR) to reduce false positives and improve robustness.

These innovations enable ISRN to achieve superior performance without relying on ImageNet pre‑training, while maintaining efficiency.

The laboratory’s broader research agenda includes face recognition, facial landmark detection, anti‑spoofing, attribute recognition, and model compression, with more than ten papers published at top conferences such as CVPR 2018 and ECCV 2018, where JD AI secured first place in multiple competition tracks.

Practically, JD AI’s face‑detection and recognition technologies have been deployed in JD‑Deqing smart stores, JD 7Fresh offline supermarkets, multimodal service robots, and intelligent commercial displays, demonstrating strong real‑world impact.

Dr. Zhou Bowen, Vice President of JD Group and head of the AI Platform and Research Division, emphasized the company’s commitment to applying cutting‑edge AI to tangible applications, and announced that the lab will host the IEEE ICME 2019 Grand Challenge on 106‑point facial landmark localization.

References: [1] https://arxiv.org/abs/1901.06651 [2] http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/WiderFace_Results.html