MogFace: A High‑Performance Face Detector with Dynamic Label Assignment, FP Context Analysis, and Pyramid‑Level Supervision

Face detection locates faces in images or video sequences and provides bounding‑box coordinates, serving as the foundation for downstream tasks such as landmark detection, attribute analysis, and recognition. The authors introduce MogFace, a high‑performance detector built on three key modules: dynamic label assignment, false‑positive (FP) context analysis, and pyramid‑layer ground‑truth (GT) assignment.

Background : Traditional label‑assignment methods rely solely on offline information (IoU, center‑point distance) or overly trust online predictions, leading to under‑utilized negative anchors, unstable assignments, and heavy hyper‑parameter tuning.

Observations :

Dynamic label assignment suffers from three problems: (1) static offline information leaves strong negative anchors unused; (2) excessive reliance on noisy online information causes erroneous assignments; (3) many hyper‑parameters make the method sensitive to dataset shifts.

FP context analysis reveals that the same false positive can become a true positive when its surrounding context changes, indicating the need for context‑aware discrimination.

Pyramid‑layer GT assignment shows that increasing GT count on a pyramid level does not always improve performance; optimal GT distribution per layer must be controlled.

Proposed Methods :

AdaptiveOnlineIncrementalAnchorMiningStrategy (Ali‑AMS) : An adaptive online incremental anchor mining approach that enhances standard anchor matching by dynamically allocating outlier faces to suitable anchors.

HierarchicalContext‑AwareModule (HCAM) : A two‑step module that explicitly encodes contextual information to differentiate FP from true positives, significantly reducing false positives.

SelectiveScaleEnhancementStrategy (SSE) : Controls the number of GTs assigned to each pyramid layer to maximize layer‑specific performance, rather than indiscriminately increasing GT density.

Experiments :

Ablation studies demonstrate the contribution of each module (Ali‑AMS, HCAM, SSE) to overall performance.

Comparison with state‑of‑the‑art methods on the WIDER FACE leaderboard shows MogFace achieving six first‑place rankings, maintaining top performance for nearly two years.

In summary, MogFace integrates dynamic label assignment, context‑aware false‑positive analysis, and pyramid‑level supervision to build a robust, high‑accuracy face detector, with extensive empirical evidence supporting its superiority over existing approaches.