POPoS: Improving Efficient and Robust Facial Landmark Detection with Parallel Optimal Position Search

Chong-Yang Xiang, Jun-Yan He, Zhi-Qi Cheng, Xiao Wu, Xian-Sheng Hua

Introduction

Facial Landmark Detection (FLD) is a fundamental task in computer vision, yet it remains constrained by a "Resolution-Efficiency Trade-off." High-performing models typically rely on high-resolution heatmaps to achieve precision, which incurs significant computational costs. Conversely, reducing heatmap resolution to save speed introduces quantization error, where the true landmark position is lost between discrete grid points.

Overview of the POPoS pipeline and the quantization error problem.

Methodology

The POPoS (Parallel Optimal Position Search) framework addresses these challenges through three primary innovations designed to achieve sub-pixel accuracy on extremely low-resolution heatmaps.

1. Pseudo-range Multilateration

Inspired by Global Positioning System (GPS) technology, POPoS treats the top $K$ heatmap peaks as "anchors" (analogous to satellites). By calculating the distances from these anchors to the landmark, the system triangulates a continuous coordinate. This redundancy ensures robustness; even if one anchor is noisy or ambiguous, the others help maintain prediction accuracy.

2. Multilateration Anchor (MA) Loss

To support the triangulation process, the authors introduced a specialized loss function. Unlike standard Mean Squared Error (MSE) which operates across the entire heatmap, MA Loss prioritizes accuracy at the specific anchor points used for decoding, ensuring high-quality input for the multilateration equations.

3. Parallel Optimal Position Search (PPPSC)

Refining landmark positions usually requires iterative optimization methods like Gauss-Newton, which are computationally expensive and poorly suited for real-time mobile deployment. POPoS replaces these with a hardware-aware "sampling and search" approach.

The parallel sampling and error computing mechanism.

As illustrated above, the PPPSC mechanism generates a grid of candidate points around an initial estimate. It then evaluates the error for all candidates simultaneously using parallel matrix operations on the GPU. This shift from sequential loops to parallel computing allows for massive speedups.

Experimental Results

The framework was evaluated on benchmark datasets including 300W, WFLW, and AFLW.

  • Extreme Efficiency: POPoS achieves a decoding speed of 1301 FPS, significantly outperforming traditional iterative methods (2.5 FPS) and previous state-of-the-art methods like KeyPosS.
  • High Accuracy at Low Resolution: On the 300W dataset using $4 \times 4$ heatmaps, the MA loss reduced prediction errors by 47% to 58% compared to baseline methods.
  • Hardware Optimization: By converting the search problem into a matrix operation, POPoS maximizes GPU utilization, making it ideal for real-time applications on edge devices.

Conclusion

POPoS demonstrates that high-resolution heatmaps are not a prerequisite for high-precision facial landmark detection. By leveraging parallel sampling and GPS-inspired multilateration, the framework provides a robust, real-time solution that overcomes the quantization errors inherent in efficient, low-resolution models. Future research may explore applying this parallel search logic to 3D facial modeling and human pose estimation.

Summary updated

Abstract

Achieving a balance between accuracy and efficiency is a critical challenge in facial landmark detection (FLD). This paper introduces Parallel Optimal Position Search (POPoS), a high-precision encoding-decoding framework designed to address the limitations of traditional FLD methods. POPoS employs three key contributions: (1) Pseudo-range multilateration is utilized to correct heatmap errors, improving landmark localization accuracy. By integrating multiple anchor points, it reduces the impact of individual heatmap inaccuracies, leading to robust overall positioning. (2) To enhance the pseudo-range accuracy of selected anchor points, a new loss function, named multilateration anchor loss, is proposed. This loss function enhances the accuracy of the distance map, mitigates the risk of local optima, and ensures optimal solutions. (3) A single-step parallel computation algorithm is introduced, boosting computational efficiency and reducing processing time. Extensive evaluations across five benchmark datasets demonstrate that POPoS consistently outperforms existing methods, particularly excelling in low-resolution heatmaps scenarios with minimal computational overhead. These advantages make POPoS as a highly efficient and accurate tool for FLD, with broad applicability in real-world scenarios.
BibTeX
@inproceedings{xiang2025popos,
  author = {Chong-Yang Xiang and Jun-Yan He and Zhi-Qi Cheng and Xiao Wu and Xian-Sheng Hua},
  title = {{POPoS: Improving Efficient and Robust Facial Landmark Detection with Parallel Optimal Position Search}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i8.32929},
}
Figure