Johns Hopkins University researchers have developed a face recognition model that routes different parts of an image through specialist neural-network components to improve matching from low-resolution footage.
Kartik Narayan and Vishal M. Patel presented FaceMoE at the European Conference on Computer Vision on September 12. The model adapts a mixture-of-experts transformer architecture for situations in which a low-resolution probe image or video must be matched against a higher-quality gallery photograph.
A conventional transformer uses a single feed-forward network after its attention operation. FaceMoE substitutes multiple feed-forward experts and a top-k router, which sends each image patch to a subset of those experts. During training, different experts learn to focus on regions such as facial contours, eyes, the nose or the mouth and adjust their contribution according to image resolution.
The sparse routing also limits the number of expert modules updated during fine-tuning. The researchers designed that mechanism to reduce catastrophic forgetting, in which a model adapted to degraded images loses recognition performance learned from high-resolution training data.
FaceMoE was evaluated on 11 datasets covering high-resolution, mixed-quality and low-resolution images. On one BRIAR protocol, it recorded true accept rates of 42.36 percent, 61.47 percent and 81.27 percent at false accept rates of 0.01 percent, 0.1 percent and 1 percent, respectively. The corresponding 1-percent result for the next-highest system in the comparison was 75.43 percent.
The model also achieved 76.18 percent rank-one retrieval accuracy on TinyFace and 44.81 percent on the surveillance-focused IJB-S benchmark. The paper says its expanded model capacity requires 1.66 times the floating-point operations of the base transformer rather than increasing computation in direct proportion to capacity.
The authors’ demographic analysis found broadly similar group-level results to the baseline, but they noted that the WebFace4M training data is weighted toward Western, young and light-skinned subjects. The model does not use balanced sampling, demographic-specific experts or a dedicated debiasing loss.
Low-resolution recognition is particularly relevant to video surveillance systems that add facial matching to existing cameras. It also complements work on training data, including an IJCB benchmark showing gains from synthetic face images. The Johns Hopkins team has published FaceMoE’s code for further evaluation.
Sources: ECCV 2026, FaceMoE paper
–
By the ID Tech Editorial Team
