Analysis of ICCV 2019 Lightweight Face Recognition Challenge Champion Solutions

Imperial College London, iQIYI, DeepGlint and others organized the Lightweight Face Recognition Challenge (LFR) at ICCV 2019, attracting 292 teams worldwide.

The challenge defined four protocols: Protocol‑1 (DeepGlint‑Light) and Protocol‑2 (DeepGlint‑Large) for image‑based face recognition, and Protocol‑3 (iQIYI‑Light) and Protocol‑4 (iQIYI‑Large) for video‑based face recognition. Each protocol imposed strict limits on FLOPs (≤1 GFlops for Light, ≤30 GFlops for Large), model size (<20 MB for Light), data type (float32) and feature dimension (512). The evaluation metric was false‑positive rate at a given threshold (FPR@1e‑8 or FPR@1e‑4).

The competition emphasized social value: developing lightweight yet high‑accuracy models that can be deployed in unrestricted dynamic surveillance video scenarios, thereby advancing research and practical translation of face‑recognition technology.

iQIYI‑Large Champion – Team “Trojans” (Hong Kong Chinese University & SenseTime X‑Lab). The team consisted of Liu Yu, Song Guanglu, Liu Jihao, Zhang Manyuan, Zhou Yucun and advisor Yan Junjie. Their solution combined a backbone network searched near 30 GFlops (similar to MNasNet/EfficientNet) with a novel loss that improved single‑model performance by 0.8 points, and a discriminative‑distribution‑based frame‑fusion strategy where fusion weights are predicted by the backbone. The code is released on GitHub: https://github.com/sciencefans/trojans-face-recognizer . Limitations include a relatively small search space, lack of data augmentation and domain‑gap handling.

DeepGlint‑Large Champion – Team from CBSR group (Automation Institute) and Winsense. Members: Liu Hao (PhD), Zhu Xiangyu, Lei Zhen, Li Ziqing, Zhang Fan and Yi Dong. They built two models (ResNet‑152 and AttentionNet‑152) with 29.5 GFLOPs, and used CosFace loss with carefully tuned margin. Their approach relied on existing components without proposing new algorithms; they noted the potential of AutoML for future improvements.

DeepGlint‑Light Champion – Team from Horizon Robotics (three researchers/engineers). Their work focused on training‑strategy exploration, network‑structure tuning and Knowledge Distillation (KD). The backbone was based on VarGNet, modified embedding/head settings and block stacking to satisfy the ≤1 GFlops constraint, followed by KD for fine‑tuning. Weaknesses include a straightforward design lacking many tricks and limited exploration of NAS or advanced KD techniques.

Overall, the LFR challenge highlighted the trade‑off between computational budget and recognition accuracy, and the winning solutions demonstrated effective architecture search, loss engineering, and frame‑fusion strategies within strict resource limits.

Source: 雷锋网