Research map / 3D vision / Human pose estimation
Alignist: CAD-Informed Orientation Distribution Estimation by Fusing Shape and Correspondences
Shishir Reddy Vutukur, Junwen Huang, Rasmus Laurvig Haugaard, Benjamin Busam, Tolga Birdal
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 |
|---|---|---|---|
| add CAD-model supervision (Alignist) when training on only 10k images | Alignist on SYMSOL-I · SYMSOL-I | log-likelihood: 10.64 -> 9.69 (only 0.95 drop with 10k images vs NF 5.06) | Improved |
| use cube-vertex positional encoding instead of IPDF PE (ablation) | Alignist · SYMSOL-I | log-likelihood: 10.09 -> 10.2 (IPDF PE vs cube PE, grid-5) | Improved |
| use Wigner matrix encoding instead of IPDF PE (ablation) | Alignist · SYMSOL-I | log-likelihood: 10.09 -> 8.59 | Hurt |
| mode-focused grid sampling instead of random sampling (ablation) | Alignist cube PE · SYMSOL-I | log-likelihood: 6.94 -> 10.2 (random vs grid-5) | Improved |
| increase sampling grid resolution from grid-5 to grid-6 (ablation) | Alignist · SYMSOL-I | log-likelihood: 10.2 -> 10.48 (L1 loss) | Improved |
| combine SDF shape expert and SurfEmb feature expert (ablation) | Alignist · SYMSOL-I | log-likelihood: 10.32 (S) / 10.18 (F) -> 10.648 (F+S) | Improved |
| replace L1 loss with generalized KL divergence (ablation) | Alignist · SYMSOL-I | log-likelihood: 10.48 -> 10.648 | Improved |
| replace L1 loss with GKL (ablation) | Alignist · T-Less | log-likelihood: 13.8 -> 14.53 | Improved |
| train SurfEmb on PBR plus real data instead of PBR only (ablation) | Alignist · T-Less | log-likelihood: 13.6 -> 13.8 | Improved |
| rely on SurfEmb expert only because SDF expert is uniform on the Sphere object | Alignist · SYMSOL-II (SphereX) | log-likelihood: 10.93 vs NF 12.37 | Hurt |
| use Alignist with CAD prior versus NF at full data | Alignist vs Normalizing Flow · SYMSOL-II | log-likelihood: 11.83 vs 12.94 | Hurt |
| longer training of Normalizing Flow (100k to 900k iterations) | Normalizing Flow · SYMSOL-I | log-likelihood: avg 5.06 -> 5.82 (10k images); icosa -2.14 -> -6.03 | Mixed |
Closest papers by meaning
- Distributional Estimation of 3D Object Orientation (NeurIPS 2026 · similarity 0.87)
- Towards Symmetry-sensitive Pose Estimation: A Rotation Representation for Symmetric Object Classes (IJCV 2026 · similarity 0.87)
- Learning Local RGB-to-CAD Correspondences for Object Pose Estimation (ICCV 2019 · similarity 0.86)
- Symmetry-Robust 3D Orientation Estimation (ICML 2025 · similarity 0.85)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.85)
- FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects (CVPR 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.
