Source-linked AI summary
Face Attention Network: An Effective Face Detector for the Occluded Faces
Jianfeng Wang, Ye Yuan, Gang Yu
TL;DR
Face detection remains challenging for faces occluded by masks, sunglasses, or other faces because higher recall can bring more false positives. FAN introduces anchor-level attention with face-specific anchor assignment and augmentation, and significantly outperforms state-of-the-art detectors on WiderFace and MAFA, especially for occluded cases.
Problem
Occluded face detection remains challenging, and improving recall can increase false positives.
Method
FAN combines a one-shot detector with face-specific anchor assignment, anchor-level attention, and data augmentation that generates cropped occluded training samples.
Results
FAN significantly outperforms state-of-the-art detectors on WiderFace and MAFA, especially in occluded cases.
Takeaways & Limitations
Anchor-level attention highlights facial-region features while relieving the risk of false positives in occluded face detection.
Abstract
from arXiv · showhide
The performance of face detection has been largely improved with the development of convolutional neural network. However, the occlusion issue due to mask and sunglasses, is still a challenging problem. The improvement on the recall of these occluded cases usually brings the risk of high false positives. In this paper, we present a novel face detector called Face Attention Network (FAN), which can significantly improve the recall of the face detection problem in the occluded case without compromising the speed. More specifically, we propose a new anchor-level attention, which will highlight the features from the face region. Integrated with our anchor assign strategy and data augmentation techniques, we obtain state-of-art results on public face detection benchmarks like WiderFace and MAFA. The code will be released for reproduction.
1. Introduction
Face detection remains difficult under occlusion because improving recall can increase false positives. FAN addresses this trade-off with anchor-level attention, specialized anchors, and augmented occluded training examples, outperforming state-of-the-art detectors on major benchmarks.
- Occlusion from masks, sunglasses, and other faces remains a challenging real-world face-detection problem.
- Recognizing an occluded face from limited visible regions can confuse the detector with similar-looking non-face regions, increasing false positives.
- FAN combines a one-shot detection pipeline, face-specific anchors, anchor-level attention, and random-crop augmentation for occluded training samples.
- Anchor-level attention is proposed to address occlusion in face detection by emphasizing features from relevant face regions.
- The reproduced one-shot RetinaNet baseline provides comparable performance with fast computation speed.
- FAN significantly outperforms state-of-the-art detectors on WiderFace and MAFA, especially for occluded cases.
2. Related Work
Related work includes single-stage and two-stage face detectors that improve detection through cascades, joint alignment, direct localization, scale handling, focal loss, context, and 3D face modeling.
- Earlier methods: Earlier face detectors used weak features such as HOG or Haar-like features in methods including Viola-Jones and deformable part models.
- Detection frameworks: The related detection literature includes SSD, RetinaNet, and two-stage Faster R-CNN detectors.
- Single-stage methods: Single-stage methods explored cascades, joint detection and landmark alignment, direct confidence and box prediction, IoU optimization, scale handling, and focal loss.
- Two-stage methods: Two-stage methods enhanced Faster R-CNN with body context, 3D face models, spatial transformations, hard-example mining, and grid loss.
3. Face Attention Network (FAN)
FAN adapts a RetinaNet-style feature-pyramid detector for faces with different scales, adding anchor-level attention and augmentation to emphasize face regions and address occlusion. Its multi-task objective jointly trains classification, box regression, and attention maps.
- 3.1. Base Framework: FAN adapts RetinaNet with FPN to detect faces at different scales across pyramid levels.The detector uses separate feature layers for different face scales and retains classification and regression subnets.
- 3.2.1. Anchor Setting: Its anchor design uses five detector layers, square and 1:1.5 aspect ratios, and scale statistics from WiderFace.More than 80% of faces in the training distribution have object scales between 16 and 406 pixels.
- 3.2.2. Attention Function: Anchor-level attention highlights face regions by supervising heatmaps with ground-truth boxes assigned to anchors at each feature layer.The layer-specific supervision produces hierarchical attention maps associated with the face sizes handled by each layer.
- 3.2.2. Attention Function: The attention maps undergo an exponential operation before being combined with feature maps, preserving context while emphasizing detection information.This design targets occluded faces, for which invisible regions may be unhelpful or harmful to detection.
- 3.2.3. Data Augmentation: Random square crops augment training data by generating occluded faces, alongside random flips and color jitter.The crop range is 0.3 to 1 of the original image's short edge, and only ground-truth boxes whose centers lie in the patch are retained.
4. Experiments
Experiments evaluate FAN on WiderFace and MAFA, examining anchor settings, attention, augmentation, benchmark accuracy, and inference efficiency. FAN achieves strong results on occluded-face benchmarks while retaining efficient computation.
- Datasets: FAN is evaluated on WiderFace and MAFA, including validation, test, masked-face, and occluded-face subsets.WiderFace includes easy, medium, and hard subsets; MAFA includes whole-dataset and occluded subsets.
- Anchor setting and assign: Anchor scale affects detector performance, while denser sampling can introduce many negative windows without improving final detection performance.The FAN anchor setting uses scales and aspect ratios selected for face detection, with an anchor assignment rule designed for greater coverage.
- Attention mechanism: 1.1% improvement on WiderFace hard and 2% improvement on MAFA masked subsets is reported when adding anchor-level attention.The attention mechanism enhances facial parts and is compared against the FAN baseline without attention.
- Data augmentation: Random-crop augmentation produces significant performance improvement and may also improve small-face detection by enlarging small faces.The augmentation addresses the limited proportion of occluded faces in WiderFace training data.
- WiderFace evaluation: FAN obtains the best reported WiderFace results across all subsets: 0.953, 0.942, and 0.888 on validation Easy, Medium, and Hard.On the test set, FAN reports 0.946, 0.936, and 0.885 for Easy, Medium, and Hard, respectively.
- MAFA evaluation: FAN significantly outperforms state-of-the-art detectors on the MAFA test set under standard testing with an IoU threshold of 0.5.MAFA is specifically designed for occluded-face detection, making it a direct evaluation of FAN's target setting.
- Inference time: FAN retains efficient computational speed while achieving state-of-the-art detection results.With short-side 1000 testing, FAN already outperforms several state-of-the-art detectors.
5. Conclusion
The paper concludes that FAN targets occluded-face detection by combining a single-stage base network with anchor-level attention. Experiments on WiderFace and MAFA validate the effectiveness and efficiency of the approach.
- Conclusion: FAN combines a specifically designed single-stage base network with anchor-level attention for occluded-face detection.The attention highlights facial-region features and is intended to relieve false-positive risk.
- Conclusion: FAN is validated on challenging WiderFace and MAFA benchmarks for effectiveness and efficiency.The supplied figures show qualitative results on the WiderFace and MAFA validation sets.