Research map / 3D vision / Depth and stereo
An Analysis of SVD for Deep Rotation Estimation
Jake Levinson, Carlos Esteves, Kefan Chen, Noah Snavely, Angjoo Kanazawa, Afshin Rostamizadeh, Ameesh Makadia
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 |
|---|---|---|---|
| replace 6D Gram-Schmidt projection with SVD orthogonalization (SVD-Train) | rotation regression network · Point cloud alignment (supervised) | mean geodesic error (deg): 2.24 -> 1.63 | Improved |
| replace 6D Gram-Schmidt projection with SVD orthogonalization | rotation regression network · Self-supervised point cloud alignment | mean geodesic error (deg): 2.39 -> 1.58 | Improved |
| replace 6D with SVD-Train | rotation regression network · ModelNet10 sofa pose | mean geodesic error (deg): 20.25 -> 18.01 | Improved |
| replace 6D with SVD-Train | rotation regression network · Human pose inverse kinematics | mean joint error (cm): 1.70 -> 1.61 | Improved |
| apply SVD projection during training (SVD-Train) rather than only at inference (SVD-Inf) (ablation) | SVD orthogonalization · Point cloud alignment (supervised) | mean geodesic error (deg): 2.64 -> 1.63 | Improved |
| apply SVD projection during training rather than only at inference (ablation) | SVD orthogonalization · ModelNet10 chair pose | mean geodesic error (deg): 21.38 -> 21.25 | No clear effect |
| apply SVD projection during training rather than only at inference (ablation) | SVD orthogonalization · Pascal3D+ mean over 12 categories | accuracy@10deg (%): 39.9 -> 35.1 | Hurt |
| replace SVD-Train with S2-Reg on Pascal3D+ (reverse comparison) | SVD-Train · Pascal3D+ mean over 12 categories | median error (deg): 11.3 vs 14.6 | Hurt |
| use SVD-Train instead of Euler angles | KITTI depth/ego-motion model · KITTI depth | Sq Rel: 3.163 -> 2.517 | Improved |
| use SVD-Train instead of axis-angle | KITTI depth model · KITTI depth | Abs Rel: 0.208 -> 0.209 | Mixed |
| train with geodesic loss instead of squared Frobenius loss (ablation) | SVD-Train · Point cloud alignment | mean geodesic error (deg): 1.63 -> 2.05 | No clear effect |
| add learning rate decay (ablation) | point cloud alignment training · Point cloud alignment | mean geodesic error (deg) | No clear effect |
Closest papers by meaning
- 3D Equivariant Pose Regression via Direct Wigner-D Harmonics Prediction (NeurIPS 2024 · similarity 0.86)
- Towards Symmetry-sensitive Pose Estimation: A Rotation Representation for Symmetric Object Classes (IJCV 2026 · similarity 0.86)
- Special Unitary Parameterized Estimators of Rotation (ICLR 2026 · similarity 0.85)
- Learning with 3D rotations, a hitchhiker's guide to SO(3) (ICML 2024 · similarity 0.85)
- CRIN: Rotation-Invariant Point Cloud Analysis and Rotation Estimation via Centrifugal Reference Frame (AAAI 2023 · similarity 0.85)
- CORA: Per-Slice Coherent Orthogonal Rotation for SVD-based Low-Rank Adaptation (NeurIPS 2026 · 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.
