Research map / 3D vision / Human pose estimation
PCLs: Geometry-Aware Neural Reconstruction of 3D Pose With Perspective Crop Layers
Frank Yu, Mathieu Salzmann, Pascal Fua, Helge Rhodin
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace root centering with PCL and PCL_inv | MLP 2D-to-3D lifting · H3.6M, 2D GT + 3D Root GT | MPJPE (mm): 45.3 \pm 0.2 -> 40.1 \boldsymbol{\pm} 0.5 | Improved |
| Replace root centering with PCL and PCL_inv | MLP 2D-to-3D lifting · MPI-INF-3DHP, 2D GT | MPJPE (mm): 74.1 \pm 0.5 -> 50.1 \boldsymbol{\pm} 0.2 | Improved |
| Replace rectangular STN cropping with PCL | ResNet-50 image-to-3D pose model · H3.6M, image input | MPJPE (mm): 96.5 -> 94.1 | Improved |
| Replace rectangular STN cropping with PCL | ResNet-50 image-to-3D pose model · MPI-INF-3DHP, image input | MPJPE (mm): 117.7 -> 109.5 | Improved |
| Apply PCL to the temporal pose sequence | T-CNN · H3.6M, original focal length | MPJPE (mm): 47.3 \pm 0.0 -> 48.8 \pm 0.3 | Hurt |
| Use PCL with known focal length instead of unnormalized T-CNN input | T-CNN · H3.6M, changed focal length | MPJPE (mm): 72.7 \pm 0.5 -> 48.9 \pm 0.2 | Improved |
| Reduce MLP + PCL linear size from 1024 to 896 (ablation) | MLP + PCL · H3.6M validation | MPJPE (mm): 43.8 -> 45.4 | Hurt |
| Reduce MLP + PCL linear size from 1024 to 768 (ablation) | MLP + PCL · H3.6M validation | MPJPE (mm): 43.8 -> 46.5 | Hurt |
| Apply only the x-axis PCL rotation post hoc (ablation) | MLP + RC · MPI-INF-3DHP, 2D GT + 3D Root GT | MPJPE (mm): 69.4 -> 67.6 | Improved |
| Apply the x- and y-axis PCL rotations post hoc (ablation) | MLP + RC · MPI-INF-3DHP, 2D GT + 3D Root GT | MPJPE (mm): 69.4 -> 66.3 | Improved |
| Train with PCL rather than applying its rotation only after training (ablation) | MLP 2D-to-3D lifting · MPI-INF-3DHP, 2D GT + 3D Root GT | MPJPE (mm): 66.3 -> 45.6 | Improved |
| Replace rectangular STN cropping with PCL | CNN pose reconstruction · Cube Dataset, unseen test set | MPJPE (mm): 13.3 -> 8.5 | Improved |
Closest papers by meaning
- PersPose: 3D Human Pose Estimation with Perspective Encoding and Perspective Rotation (ICCV 2025 · similarity 0.87)
- Cameras as Relative Positional Encoding (NeurIPS 2025 · similarity 0.86)
- LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.85)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.85)
- Know Your Neighbors: Improving Single-View Reconstruction via Spatial Vision-Language Reasoning (CVPR 2024 · similarity 0.85)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.84)
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.
