Research map / 3D vision / Human pose estimation
Image to Sphere: Learning Equivariant Features for Efficient Pose Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Spherical (S2) equivariant pipeline with spatial orthographic projection, versus MLP-based Fourier coefficients (ablation) | I2S pose predictor · ModelNet10-SO(3) limited training views | Acc@15: 0.623 vs 0.360 | Helped |
| Spatial orthographic projection instead of Fourier-coefficient projection (ablation) | I2S feature projection · ModelNet10-SO(3) limited training views | Acc@15: 0.623 vs 0.523 (spatial filter variants) | Helped |
| Spherical convolution instead of icosahedral discrete group convolution (ablation) | I2S · ModelNet10-SO(3) limited training views | Acc@15: 0.623 vs 0.455 | Helped |
| Filter parametrization: spatial filter instead of Fourier filter (with spatial projection) (ablation) | I2S · ModelNet10-SO(3) limited training views | Acc@15: 0.610 vs 0.623 | No clear effect |
| Add final SO(3) group convolution after S2 convolution (ablation) | I2S · ModelNet10-SO(3) full training set | MedErr: 14.4 with vs 15.6 without (Acc@15 unchanged at 0.729) | Mixed |
| Add final SO(3) group convolution after S2 convolution (ablation) | I2S · ModelNet10-SO(3) limited training views | Acc@15: 0.623 vs 0.616 | Helped |
| Add a second SO(3) convolution (ablation) | I2S · ModelNet10-SO(3) full training set | MedErr: 19.0 vs 14.4 with one conv | Hurt |
| Maximum frequency L higher than 6 (L=8, 10) instead of L=6 (ablation) | I2S · ModelNet10-SO(3) limited training views | Acc@15: 0.605 (L=8), 0.599 (L=10) vs 0.623 (L=6) | Hurt |
| I2S vs best baseline (regression/distribution) on PASCAL3D+ | I2S · PASCAL3D+ | avg. median rotation error: 9.8 vs 10.1 | Helped |
| I2S vs implicit baseline on SYMSOL II with 100k images | I2S · SYMSOL I and II | avg. log likelihood: 3.41 vs 4.10 (SYMSOL I); 4.84 vs 7.57 (SYMSOL II) | Hurt |
| Reduce SYMSOL training from 100k to 10k images per shape with I2S | I2S · SYMSOL I | avg. log likelihood: 2.98 (10k) vs 3.41 (100k) | Mixed |
Closest papers by meaning
- 3D Equivariant Pose Regression via Direct Wigner-D Harmonics Prediction (NeurIPS 2024 · similarity 0.92)
- Equivariant Single View Pose Prediction Via Induced and Restriction Representations (NeurIPS 2023 · similarity 0.91)
- Leveraging Equivariant Features for Absolute Pose Regression (CVPR 2022 · similarity 0.90)
- Learning to Orient Surfaces by Self-supervised Spherical CNNs (NeurIPS 2020 · similarity 0.90)
- E2PN: Efficient SE(3)-Equivariant Point Network (CVPR 2023 · similarity 0.89)
- Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation (ICLR 2025 · similarity 0.87)
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.