Abstract: Scale variation remains a significant hurdle in face detection, particularly for small or occluded faces. EMface introduces Receptive Field Pyramids (RFP), a lightweight module that allows detectors to adaptively "see" faces at multiple scales within the same feature map. By employing parallel dilated convolutions with shared weights and a branch pooling strategy, EMface achieves state-of-the-art accuracy on challenging benchmarks like WIDER FACE (Hard) while remaining significantly faster than previous heavy-duty detectors.
The Challenge of Scale Discretization
Modern face detectors typically use Feature Pyramid Networks (FPN) to handle scale variation, assigning small faces to high-resolution maps and large faces to low-resolution maps. However, this approach suffers from discretization: faces exist on a continuous scale, yet FPNs force them into discrete levels.
In a standard FPN, faces of significantly different sizes (e.g., 50x50 vs. 100x100 pixels) may be assigned to the same feature map level. Because standard convolutional layers have a fixed receptive field, a single layer might be optimized for one size but suboptimal for another. This inconsistency makes it particularly difficult to detect faces in "Hard" datasets where faces are tiny, blurry, or partially occluded.
Proposed Method: Receptive Field Pyramids (RFP)
The authors propose the RFP module to diversify the scales a network can handle without a massive increase in parameter count.
Core Components:
- Multi-Branch Dilated Convolutions: Instead of a single $3 \times 3$ convolution, the RFP module uses three parallel branches with different dilation rates ($d=1, 3, 5$). This expands the receptive field without adding parameters or losing resolution.
- Weight Sharing: To maintain efficiency, all three branches share the exact same weights. This forces the kernels to learn scale-invariant features.
- Branch Pooling: During training, branch outputs are fused via average pooling to balance representations across different receptive fields.
- The Inference Trick: During testing, the authors found they could drop the $d=1$ and $d=5$ branches, keeping only the middle $d=3$ branch. Due to the weight sharing and pooling during training, this single branch retains most of the multi-scale knowledge, drastically reducing computation time.
Experimental Results
EMface was evaluated on the WIDER FACE benchmark (Easy, Medium, and Hard subsets) and the UFDD (Unconstrained Face Detection Dataset), which focuses on extreme weather and motion blur.
Key Findings:
- Accuracy on WIDER FACE (Hard): EMface achieved 87.2% AP, outperforming heavyweights like DSFD (84.6%) and S3FD (83.4%).
- Efficiency: On a V100 GPU, EMface (ResNet50 backbone) runs at 182ms per image, nearly 3x faster than DSFD (508ms) while maintaining higher accuracy on difficult cases.
- Robustness: On the UFDD benchmark, EMface produced superior precision-recall curves, demonstrating its ability to handle real-world degradations like blur and weather better than previous models like HR-ER and SSH.
Conclusion
EMface demonstrates that a feature map's level should not dictate a single receptive field. By providing a "pyramid" of receptive fields within a single layer, the network can adaptively capture faces that fall between the cracks of standard FPN levels. While not yet "real-time" for high-resolution inputs, the RFP module offers a general-purpose architectural improvement that could be extended to general object detection and semantic segmentation.
Summary updated