MGFace boosts masked-face ID accuracy 20x faster with conditional routing

Face identification systems have achieved remarkable accuracy under normal conditions, but performance degrades sharply when faces are partially occluded—especially by facial masks. Existing re-ranking approaches that exploit patch-level similarities improve robustness but are computationally expensive, limiting their scalability. To address this, researchers from the paper introduce MGFace (Mask-Gated Face Matching via Conditional Similarity Routing). The pipeline first predicts whether a query face is masked or not. For unmasked queries, it uses standard global embedding matching, which is fast and efficient. For masked queries, it activates a mask-aware patch-level re-ranking mechanism that focuses on reliable upper-face regions while avoiding unnecessary computation. This conditional design ensures accuracy gains without sacrificing speed.

Experiments on the extended LFW-Mask dataset demonstrate MGFace's effectiveness. With the FaceNet backbone, it achieves over 80% identification accuracy; with the more powerful ArcFace backbone, accuracy exceeds 90%. Compared to a prior EMD-based re-ranking method, MGFace delivers better identification performance while being approximately 20x faster in query time. This efficiency makes it practical for large-scale retrieval scenarios, such as security checkpoints or authentication systems. The source code is provided, enabling further research and deployment. MGFace represents a pragmatic advance in handling occluded faces without heavy computational overhead, balancing accuracy and speed for real-world applications.