Abstract
In recent years, face detection has experienced significant performance improvement with the boost of deep convolutional neural networks. In this report, we reimplement the state-of-the-art detector SRN and apply some tricks proposed in the recent literatures to obtain an extremely strong face detector, named VIM-FD. In specific, we exploit more powerful backbone network like DenseNet-121, revisit the data augmentation based on data-anchor-sampling proposed in PyramidBox, and use the max-in-out label and anchor matching strategy in SFD. In addition, we also introduce the attention mechanism to provide additional supervision. Over the most popular and challenging face detection benchmark, i.e., WIDER FACE, the proposed VIM-FD achieves state-of-the-art performance.
Introduction
While deep learning has revolutionized face detection, the task remains challenging due to extreme variations in scale, pose, occlusion, and lighting. Most modern detectors are categorized as either one-stage (fast but less accurate) or two-stage (accurate but slower). VIM-FD (Robust and High Performance Face Detector) bridges this gap by "organically integrating" several powerful techniques from recent literature into a cohesive, high-performance system.
The primary goal of VIM-FD is to excel at detecting faces across extreme scales, particularly targeting the "Hard" subset of the WIDER FACE benchmark, which contains a high volume of tiny, difficult-to-detect faces. A notable demonstration of its robustness is its ability to detect 890 faces in a single crowded "world record" selfie.
Methodology
VIM-FD is built upon the Selective Refinement Network (SRN) as a baseline, incorporating several key architectural and training upgrades:
- DenseNet-121 Backbone: Replacing the standard ResNet-50, DenseNet-121 provides a more powerful feature extractor. It utilizes a 6-level feature pyramid to capture multi-scale information effectively.
- STC and STR Modules: The architecture employs Selective Two-step Classification (STC) to filter out easy negative samples early and Selective Two-step Regression (STR) to coarsely adjust anchor locations before final prediction.
- Anchor-Level Attention: Borrowing from the Face Attention Network (FAN), an attention subnet is applied to feature maps P2 through P7. This provides additional supervision, forcing the model to highlight facial regions and ignore background clutter.
- Max-in-out Labels: To reduce false positives among small objects, the model predicts multiple scores for background/foreground and selects the maximum, making the classifier more robust against "hard" negatives.
- Scale-Compensated Anchor Matching: To address the failure of standard anchor matching for tiny faces, the authors lowered IoU thresholds and used a scale compensation strategy to ensure small faces receive sufficient training samples.
- Data-Anchor-Sampling: This method resizes images so that faces are more likely to match the scales of the network's anchors, diversifying the scale distribution during training.
Experimental Setup
The model was evaluated on WIDER FACE, the gold standard for face detection benchmarking. The dataset contains over 32,000 images and nearly 400,000 annotated faces, divided into three difficulty levels: Easy (large faces), Medium (moderate scales), and Hard (tiny, occluded, or blurry faces). Training was conducted exclusively on the WIDER FACE training set.
Results and Discussion
VIM-FD achieved top-tier performance on the WIDER FACE benchmark. On the validation set, it reached 96.7% (Easy), 95.7% (Medium), and 90.7% (Hard). On the testing set, it achieved 96.2% (Easy), 95.3% (Medium), and 90.2% (Hard).
The results highlight several key insights:
- The Power of Integration: State-of-the-art performance can be achieved by cleverly combining existing, proven modules like attention, better backbones, and smart sampling.
- Small Face Robustness: Strategies specifically targeting small faces, such as scale-compensated matching and data-anchor-sampling, provided the most significant gains.
- Attention as Supervision: Treating attention as a supervised task helps the model learn where to look for faces, significantly reducing false positives in complex backgrounds.
Conclusion
VIM-FD demonstrates that a robust integration of modern detection "tricks" can push the boundaries of face detection performance, especially for small and crowded faces. While the model achieves impressive accuracy, future work could focus on optimizing inference speed and computational cost for real-time applications on edge devices, potentially through model pruning or knowledge distillation.
Summary updated