Abstract
Benefiting from the pioneering design of generic object detectors, significant achievements have been made in the field of face detection. Typically, the architectures of the backbone, feature pyramid layer, and detection head module within the face detector all assimilate the excellent experience from general object detectors. However, several effective methods, including label assignment and scale-level data augmentation strategy, fail to maintain consistent superiority when applying on the face detector directly. Concretely, the former strategy involves a vast body of hyper-parameters and the latter one suffers from the challenge of scale distribution bias between different detection tasks, which both limit their generalization abilities. Furthermore, in order to provide accurate face bounding boxes for facial down-stream tasks, the face detector imperatively requires the elimination of false alarms. As a result, practical solutions on label assignment, scale-level data augmentation, and reducing false alarms are necessary for advancing face detectors. In this paper, we focus on resolving three aforementioned challenges that exiting methods are difficult to finish off and present a novel face detector, termed MogFace. In our Mogface, three key components, Adaptive Online Incremental Anchor Mining Strategy, Selective Scale Enhancement Strategy and Hierarchical Context-Aware Module, are separately proposed to boost the performance of face detectors. Finally, to the best of our knowledge, our MogFace is the best face detector on the Wider Face leader-board, achieving all champions across different testing scenarios.
Introduction
While generic object detectors like Faster R-CNN or SSD have advanced significantly, face detection presents unique hurdles that standard "tricks" often fail to address. MogFace identifies three primary pain points: sub-optimal label assignment based on static rules, extreme scale variance, and the prevalence of false alarms.
Standard data augmentation, which scales images uniformly, often fails to provide the specific scale information required for the massive range of face sizes. As shown in the scale distribution analysis, the relative scale of faces in datasets like WIDER FACE is significantly more skewed towards smaller objects compared to general object datasets like COCO, necessitating a more specialized approach to scale augmentation.
Methods
The authors propose a "three-pillar" approach to build MogFace, focusing on adaptive learning and context awareness.
Adaptive Online Incremental Anchor Mining (Ali-AMS)
Instead of relying solely on static IoU rules (offline information), Ali-AMS utilizes the model's own "online" predictions during training. By treating predicted scores as a quality metric, it mines better anchors for faces that are difficult to match, ensuring high-quality supervision as the model evolves.
Selective Scale Enhancement (SSE)
The authors discovered that matching too many faces to a specific feature layer can actually degrade performance. SSE is a data augmentation strategy that intelligently controls the distribution of face scales across the model's pyramid layers to maximize learning efficiency, adhering to an "appropriate match" principle.
Hierarchical Context-Aware Module (HCAM)
To eliminate false alarms—such as face-like textures on walls or calendars—HCAM analyzes the "neighborhood" around a detection. By encoding surrounding context, the model can distinguish true positives from high-confidence mistakes that lack human-centric global context.
Results and Discussion
MogFace achieved state-of-the-art results, securing "all champions" across testing scenarios on the WIDER FACE leaderboard at the time of publication.
- WIDER FACE Hard: Ali-AMS improved Average Precision (AP) by 0.8% over the baseline.
- WIDER FACE Easy: The SSE strategy boosted AP by 1.0%.
- False Alarm Reduction: The HCAM module was highly effective, reducing the Number of False Alarms (NFA) on the validation set from 948 down to 192 while simultaneously increasing detection accuracy.
- Generalization: The model achieved near-perfect 99.2% on FDDB and state-of-the-art results on AFW and PASCAL Face.
Conclusion
MogFace demonstrates that rethinking standard object detection practices for the specific challenges of faces—particularly scale distribution and context—leads to significant performance gains. While highly effective, the SSE strategy currently relies on prior statistical analysis of the dataset. Future work will explore making these scale-distribution controls more automated and dynamic, alongside optimizing the HCAM module for real-time inference on edge devices.
Summary updated