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

TL;DR

POP0S targets the accuracy–efficiency trade-off in facial landmark detection by combining pseudo-range multilateration with a multilateration anchor loss and a GPU-friendly, single-step parallel decoding pipeline. The method uses top-K heatmap anchors to form distance equations, transforms heatmaps into sub-pixel distance maps, and solves for landmark positions via PPPSC, supported by Iterative Gauss–Newton optimization for PRM. Across five datasets, POPoS achieves lower Normalized Mean Error than many existing heatmap and coordinate-regression approaches, with particular strength at low heatmap resolutions and competitive computational efficiency. The combination of MA loss, PPPSC, and parallel computation enables real-time or near-real-time FLD with improved robustness to heatmap quantization and multi-modal predictions, making it applicable to mobile and edge devices.

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 a highly efficient and accurate tool for FLD, with broad applicability in real-world scenarios.

Figures (6)

• Figure 1: Quantization error in the coordinate decoding process. The discrepancy between the ground truth (GT) coordinates and the highest probability point in the heatmap illustrates the challenge of achieving precise localization and lossless decoding. [Best viewed in color with zoom]
• Figure 2: Parallel Optimal Position Search (POPoS) framework. The Multilateration Anchor Loss is designed during the training phase to optimize distance prediction. During model inference, POPoS samples the top-K response positions as anchor points for distance map decoding. [Best viewed in color with zoom]
• Figure 3: Keypoint detection accuracy at heatmap resolutions of $64 \times 64$, $32 \times 32$, $16 \times 16$, $8 \times 8$, and $4 \times 4$. [The smaller radar map denotes the better performance]
• Figure 4: Effect of Gauss-Newton Optimization (IGNO), Multilateration Anchor (MA) loss, and Potential Position Parallel Sampling and Computing (PPPSC).
• Figure 5: Detailed analysis of POPoS performance factors: (a) Effect of sampling area (b) Effect of sampling rate $\tau$ (c) Anchor number and (d) MA loss weight, each illustrating their impact on the accuracy of predictions.