Research map / 3D vision / Face and shape
Learning an Effective Equivariant 3D Descriptor Without Supervision
Riccardo Spezialetti, Samuele Salti, Luigi Di Stefano
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 |
|---|---|---|---|
| orient the equivariant descriptor with an external LRF (FLARE) instead of self-orienting (ablation) | equivariant spherical CNN descriptor · 3DMatch average | recall: 0.8918 -> 0.9420 | Improved |
| orient the equivariant descriptor with an external LRF instead of self-orienting (ablation) | equivariant spherical CNN descriptor · Rotated 3DMatch average | recall: 0.8778 -> 0.9387 | Improved |
| learn an equivariant descriptor then orient it, instead of learning an invariant one from PPF (PPFFoldNet) | FoldingNet-style unsupervised descriptor · 3DMatch average | recall: 0.7182 -> 0.9420 | Improved |
| learn equivariant descriptor instead of invariant CGF | learned descriptor · 3DMatch average | recall: 0.4776 -> 0.9420 | Improved |
| use Spherical CNN encoder instead of PointNet to obtain rotation invariance after rotating the output (ablation) | descriptor encoder · descriptor distance under input rotation | invariance | Improved |
| use equivariant bottleneck instead of invariant bottleneck (max pooling) (ablation) | folding decoder reconstruction · local patch reconstruction after 10K iterations | reconstruction quality | Improved |
| rotate inputs of the benchmark randomly (rotated 3DMatch) | Ours LRF · 3DMatch average | recall: 0.9420 -> 0.9387 | No clear effect |
| rotate inputs of the benchmark randomly (rotated 3DMatch) | 3DMatch descriptor · 3DMatch average | recall: 0.5726 -> 0.0113 | Hurt |
| rotate inputs of the benchmark randomly (rotated 3DMatch) | PPFNet · 3DMatch average | recall: 0.6231 -> 0.0026 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- 3DMatch — 3DMatch: Learning Local Geometric Descriptors From RGB-D Reconstructions (CVPR 2017) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unsupervised Learning of Local Equivariant Descriptors for Point Clouds (TPAMI 2021 · similarity 0.94)
- 3D Equivariant Pose Regression via Direct Wigner-D Harmonics Prediction (NeurIPS 2024 · similarity 0.88)
- Neural Isometries: Taming Transformations for Equivariant ML (NeurIPS 2024 · similarity 0.87)
- TetraSphere: A Neural Descriptor for O(3)-Invariant Point Cloud Analysis (CVPR 2024 · similarity 0.87)
- An Intuitive Multi-Frequency Feature Representation for SO(3)-Equivariant Networks (ICLR 2024 · similarity 0.87)
- Multiview Equivariance Improves 3D Correspondence Understanding with Minimal Feature Finetuning (ICLR 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.
