Research map / Architectures and efficiency / RNNs and normalization
POPoS: Improving Efficient and Robust Facial Landmark Detection with Parallel Optimal Position Search
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes POPoS, a heatmap-based facial landmark detection framework that decodes landmarks via pseudo-range multilateration over top-K distance-map anchors, trained with a multilateration anchor loss and a GPU-parallel candidate-sampling solver.
Problem
Heatmap-based facial landmark detection suffers quantization and encoding/decoding errors at low heatmap resolutions, and accurate distribution-aware decoding is slow, so accuracy and efficiency trade off poorly.
Approach
Treats heatmap-to-distance decoding as pseudo-range multilateration using the top-K heatmap peaks as anchors, adds a loss on the top-K region (MA loss) to align training with decoding, and solves the least-squares objective by brute-force parallel sampling of candidate sub-pixel positions on the GPU instead of iterative Gauss-Newton.
Key results
- POPoS at 64x64 heatmaps reports NME 3.95% on WFLW, 1.34% on AFLW, and 3.44% on COFW, with 9.7M parameters.
- POPoS decoding runs at 1301 FPS, about 15x faster than KeyPosS (96.5 FPS) and 6x faster than DarkPose (186.2 FPS), though slower than one-hot (2819 FPS).
- The MA loss reduces 4x4 heatmap prediction error by 47% on 300W and 58% on COFW.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add iterative Gauss–Newton optimization for pseudo-range multilateration (ablation) | KeyPosS baseline · 300W and COFW | NME | Helped |
| Add Multilateration Anchor loss (ablation) | POPoS · 300W, 4×4 heatmap | prediction error: 47% | Helped |
| Add Multilateration Anchor loss (ablation) | POPoS · COFW, 4×4 heatmap | prediction error: 58% | Helped |
| Add Potential Position Parallel Sampling and Computing (ablation) | IGN Optimization · Five datasets at varying resolutions | stability and performance | Helped |
| Increase sampling range beyond 1–2 pixels around the maximum (ablation) | POPoS decoding · 300W, 32×32 heatmap | decoding accuracy | No clear effect |
| Increase sampling rate beyond approximately 10 sampling points per pixel length (ablation) | POPoS decoding · 300W, 32×32 heatmap | accuracy | No clear effect |
| Use too few anchors (ablation) | POPoS decoding · 32×32 heatmap | prediction accuracy | Hurt |
| Use too many anchors (ablation) | POPoS decoding · 32×32 heatmap | calculation error | Hurt |
| Add local anchor loss, with weight ratio 6 (ablation) | POPoS · 300W | prediction accuracy | Helped |
| Use POPoS instead of KeyPosS | Facial landmark decoding · Decoding methods | FPS: 96.5 -> 1301 | Helped |
| Use POPoS instead of one-hot decoding | Facial landmark decoding · Decoding methods | FPS: 2819 -> 1301 | Hurt |
Surprising findings
- Increasing the sampling range beyond 1-2 pixels around the maximum did not improve decoding accuracy, which the authors attribute to encoding-stage errors that decoding cannot fix.
- The IGNO solver is very slow (2.5 FPS, 405 ms per decode), which motivates the parallel PPPSC alternative.
Limitations
- Gains at higher resolutions and on 300W are modest; 300W NME of 3.28% is not best among compared methods (e.g., STAR Loss 2.87%).
- Decoding is still roughly 2x slower than one-hot decoding.
- Hyperparameters (sampling range, anchor count, MA weight) are tuned on 300W and then applied to other datasets.
Experimental setup
Datasets: COFW AFLW COCO-WholeBody 300W WFLW
Benchmarks: WFLW 300W AFLW COFW COCO-WholeBody
Metrics: NME FPS GFlops parameter count
Baselines: One-hot decoding Two-hot DarkPose KeyPosS AWing STAR Loss PIPNet SLPT DTLD RePFormer
Builds on: HRNet KeyPosS DarkPose MMPose
Closest papers by meaning
- AnchorFace: An Anchor-based Facial Landmark Detector Across Large Poses (AAAI 2021 · similarity 0.87)
- Improving Landmark Localization With Semi-Supervised Learning (CVPR 2018 · similarity 0.85)
- Super-FAN: Integrated Facial Landmark Localization and Super-Resolution of Real-World Low Resolution Faces in Arbitrary Poses With GANs (CVPR 2018 · similarity 0.85)
- Look at Boundary: A Boundary-Aware Face Alignment Algorithm (CVPR 2018 · similarity 0.85)
- T-FAKE: Synthesizing Thermal Images for Facial Landmarking (CVPR 2025 · similarity 0.83)
- FaceLift: Semi-supervised 3D Facial Landmark Localization (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.