Research map / Point clouds / Segmentation and detection
Formula-Supervised Visual-Geometric Pre-training
Ryosuke Yamada, Kensho Hara, Hirokatsu Kataoka, Koshi Makihara, Nakamasa Inoue, Rio Yokota, Yutaka Satoh
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 |
|---|---|---|---|
| pre-train on both fractal images and point clouds (V+G) instead of visual only (ablation) | FSVGP VG-FractalDB-1k · Image CLS average (7 datasets) | accuracy: 89.3 -> 90.0 | Improved |
| pre-train on both modalities (V+G) instead of geometric only (ablation) | FSVGP VG-FractalDB-1k · Image CLS average (7 datasets) | accuracy: 67.2 -> 90.0 | Improved |
| add geometric modality to visual pre-training (ablation) | FSVGP VG-FractalDB-1k · 3D CLS average (ModelNet40, ScanObjectNN) | accuracy: 88.0 -> 88.3 | Improved |
| use formula-supervised consistency label instead of MAE self-supervision (ablation) | VG-FractalDB-1k pre-training · ImageNet100 | accuracy: 80.3 -> 92.0 | Improved |
| use formula-supervised consistency label instead of MAE self-supervision (ablation) | VG-FractalDB-1k pre-training · ModelNet40 | accuracy: 92.8 -> 92.9 | Improved |
| use fractal generation instead of Perlin noise (ablation) | VG-FractalDB · ImageNet100 | accuracy: 90.7 -> 92.0 | Improved |
| use fractal generation instead of Perlin noise (ablation) | VG-FractalDB · ModelNet40 | accuracy: 92.6 -> 92.9 | Improved |
| use fractal data instead of ShapeNet CAD projections at equal 50k size (ablation) | VG-FractalDB · ImageNet100 | accuracy: 87.3 -> 87.9 | Improved |
| use fractal data instead of ShapeNet CAD projections at equal 50k size (ablation) | VG-FractalDB · ModelNet40 | accuracy: 92.7 -> 92.8 | Improved |
| FSVGP with VG-FractalDB-1k versus VisualAtom-21k with 21x more data | FSVGP · Image CLS average (7 datasets) | accuracy: 89.7 -> 90.0 | Improved |
| FSVGP versus real-data SSL pre-training (MAE, DINO) and supervised ImageNet | FSVGP · Image CLS average (7 datasets) | accuracy: 90.0 vs 92.5 (MAE) | Hurt |
| FSVGP versus ShapeNet SSL methods for 3D classification | FSVGP · 3D CLS average (ModelNet40, ScanObjectNN) | accuracy: 88.5 vs 89.5 (Point-BERT) | Hurt |
Closest papers by meaning
- Visual Atoms: Pre-Training Vision Transformers With Sinusoidal Waves (CVPR 2023 · similarity 0.88)
- Replacing Labeled Real-Image Datasets With Auto-Generated Contours (CVPR 2022 · similarity 0.87)
- Self-Supervised Pretraining of 3D Features on Any Point-Cloud (ICCV 2021 · similarity 0.87)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.85)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.85)
- Primitive Vision: Improving Diagram Understanding in MLLMs (ICML 2025 · similarity 0.85)
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.
