Abstract
Facial landmark alignment is critical for face recognition and animation, yet traditional top-down methods struggle with scalability in multi-face scenarios. This paper introduces FoxNet, a bottom-up architecture that detects all landmarks for all faces in a single pass. By employing a "Fox Block" for multiscale feature extraction and a Cosine Discriminative Loss for landmark clustering, FoxNet achieves near-constant inference speed regardless of the number of faces. Experimental results demonstrate that FoxNet maintains high precision and robustness to occlusions, offering a significant performance advantage for real-time applications.
Introduction
Traditional facial landmark alignment typically follows a top-down approach: first detecting individual faces and then running a landmark detector on each cropped face. This results in a linear bottleneck ($O(n)$ complexity), where processing time increases significantly with the number of faces. Additionally, these methods often rely on Non-Maximum Suppression (NMS), which can struggle with overlapping faces and occlusions.
FoxNet addresses these challenges by shifting to a bottom-up paradigm. Instead of sequential detection and alignment, it treats the problem as a simultaneous pixel-embedding and clustering task, allowing for efficient processing of images containing numerous individuals.
Methodology
The FoxNet architecture is built upon a ResNet50 backbone that branches into two simultaneous outputs: a Segmentation Branch for identifying landmark candidates and a Feature Branch for generating high-dimensional pixel embeddings.
The Fox Block and Cosine Discriminative Loss
A key innovation is the Fox Block, designed to capture multiscale global context. Unlike human pose estimation, facial landmarks lack explicit physical "connections" (like limbs). The Fox Block uses four different kernel sizes (1x1, 3x3, 5x5, and 7x7) for average pooling to help the model associate landmarks with the correct face.
To group these landmarks, FoxNet utilizes a Cosine Discriminative Loss. Unlike standard Euclidean-based losses, this function forces embeddings onto a hypersphere, maximizing the distance between different faces while minimizing the distance within the same face. During inference, the Mean Shift algorithm clusters these embeddings to finalize the alignment.
Key Results
The primary advantage of FoxNet is its constant-time inference. While top-down methods like SSH+LAB show a steep linear increase in runtime as more faces are added, FoxNet's CNN processing time remains stable.
Quantitatively, FoxNet processes images in approximately 51.50 ms, with a marginal increase of only 2.06 ms per additional face. In contrast, traditional methods add roughly 71 ms per person. On the WFLW dataset, FoxNet achieved a 5.80% Normalized Mean Error (NME). While slightly behind the state-of-the-art LAB (5.27%) in clean conditions, it outperformed LAB in "Occlusion" and "Blur" categories, highlighting the benefits of its global receptive field.
Conclusion and Future Work
FoxNet represents a significant step forward for real-time multi-face alignment. By moving to a bottom-up approach, it effectively "breaks the speed limit" imposed by traditional top-down pipelines. While there is still a small accuracy gap compared to the most complex single-face models, the robustness and efficiency of FoxNet make it a powerful tool for crowded environments. Future research will focus on incorporating explicit geometric constraints to further refine landmark precision.
Summary updated