Research map / 3D vision / Face and shape
Geometric Correspondence Fields: Learned Differentiable Rendering for 3D Pose Refinement in the Wild
Alexander Grabner, Yaming Wang, Peizhao Zhang, Peihong Guo, Tong Xiao, Peter Vajda, Peter M. Roth, Vincent Lepetit
Read the paperCheck 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 |
|---|---|---|---|
| use a larger receptive field (7x7 kernels) (ablation) | correspondence branch · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.75 -> 1.34 | Improved |
| use more layers (4 instead of 2) in correspondence branch (ablation) | correspondence branch · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.58 -> 1.34 | Improved |
| use deeper mapping branches instead of a single Conv-BN-ReLU block (ablation) | mapping branch · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.55 -> 1.34 | Improved |
| add geometric attention (ablation) | Our Refinement · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.45 -> 1.34 | Improved |
| concatenate depth, normal and object coordinate renderings instead of mask only (ablation) | Our Refinement · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.44 -> 1.34 | Improved |
| concatenate three renderings instead of depth only (ablation) | Our Refinement · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.38 -> 1.34 | Improved |
| concatenate three renderings instead of normal only (ablation) | Our Refinement · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.36 -> 1.34 | Improved |
| replace Mask Refinement with learned correspondence-field refinement | Mask Refinement · Pix3D (GT models) MedErr_R,t | MedErr_R,t (x1e-2): 2.96 -> 1.34 | Improved |
| replace Image Refinement with learned correspondence-field refinement | Image Refinement · Pix3D (GT models) MedErr_R,t | MedErr_R,t (x1e-2): 4.31 -> 1.34 | Improved |
| use retrieved 3D models instead of ground-truth models | Our Refinement · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 1.34 -> 2.14 | Hurt |
| add refinement with retrieval on top of feed-forward pose baseline | Baseline · Pix3D MedErr_R,t | MedErr_R,t (x1e-2): 4.76 -> 2.14 | Improved |
| run Image Refinement for 1000 iterations instead of 20 | Image Refinement · Pix3D | MedErr: 20 iterations best vs 1000 | Hurt |
Closest papers by meaning
- Soft Rasterizer: A Differentiable Renderer for Image-Based 3D Reasoning (ICCV 2019 · similarity 0.89)
- A General Differentiable Mesh Renderer for Image-Based 3D Reasoning (TPAMI 2020 · similarity 0.89)
- Differentiable Volumetric Rendering: Learning Implicit 3D Representations Without 3D Supervision (CVPR 2020 · similarity 0.88)
- Unifying Correspondence Pose and NeRF for Generalized Pose-Free Novel View Synthesis (CVPR 2024 · similarity 0.87)
- Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level Object Pose Estimation (ICLR 2024 · similarity 0.87)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.86)
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.
