Didi AI Labs' DFS Face Detection Algorithm Achieves Top Rankings on the WIDER FACE Benchmark

The latest evaluation of the world‑most authoritative face detection benchmark, WIDER FACE, shows that the DFS algorithm jointly developed by Didi AI Labs and the PRIS team of Beijing University of Posts and Telecommunications achieved five first‑place and one second‑place results across the Easy, Medium, and Hard subsets.

WIDER FACE, created by the Chinese University of Hong Kong in 2016, is one of the largest and most challenging public face detection datasets, containing 32,203 images with 393,703 annotated faces. The dataset is split into 40% training, 10% validation, and 50% testing, and each split is further categorized by difficulty (Easy, Medium, Hard).

In the official validation and test sets, the DFS algorithm obtained AP scores of 96.3% (Easy), 95.4% (Medium), and 90.7% (Hard), surpassing many leading companies and academic institutions.

The DFS method introduces a Feature Fusion Pyramid (FFP) that merges high‑level semantic information with low‑level detailed features using spatial and channel attention mechanisms. This design preserves fine details while enriching low‑level features with contextual cues.

Additionally, DFS incorporates a semantic‑segmentation branch that provides auxiliary supervision for a single‑stage detector. The branch supplies multi‑scale segmentation signals, encouraging the network to learn attention maps and better feature representations for faces of various sizes.

These innovations have been deployed in Didi’s platform for driver‑identity verification, in‑vehicle video privacy protection, and other real‑time perception scenarios, delivering mature solutions for external partners.

Didi’s computer‑vision team has a history of top‑ranked results in international challenges, such as the CVPR 2018 WAD autonomous‑driving contest and the ECCV 2018 COCO & Mapillary challenge, where they secured world‑first positions.

The continued success of the DFS algorithm underscores Didi’s strong capabilities in AI and computer vision, and the company plans to further explore and optimize perception, interaction, and intelligent transportation technologies.

References: [1] WIDER FACE [2] DFS Algorithm: Learning Better Features for Face Detection with Feature Fusion and Segmentation Supervision