Research map / 3D vision / 3D generation
ReGen3D: Generalizable Unified Representation Learning for 3D Understanding
Haotian Zhang, Jincen Jiang, Yuhang Li, Jian Jun Zhang, Meili Wang, Jian Chang
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 URT (unified tokenization) to the baseline (ablation) | ReGen3D baseline · Voxel held out, classification | accuracy (%): 36.2 -> 51.8 | Improved |
| add URT (unified tokenization) to the baseline (ablation) | ReGen3D baseline · Voxel held out, segmentation | mIoU (%): 22.1 -> 52.4 | Improved |
| add URT to the baseline (ablation) | ReGen3D baseline · Voxel held out, reconstruction | CD (x10^-3): 57.2 -> 55.7 | No clear effect |
| add SMI (shared memory interaction) on top of URT (ablation) | ReGen3D · Voxel held out, classification | accuracy (%): 51.8 -> 72.6 | Improved |
| add SMI on top of URT (ablation) | ReGen3D · Voxel held out, segmentation | mIoU (%): 52.4 -> 58.6 | Improved |
| add SMI on top of URT (ablation) | ReGen3D · Training representations P/M/G reconstruction | CD (x10^-3): 27.2 -> 23.4 (Point) | Improved |
| add ERA (efficient representation adapter) on top of URT and SMI (ablation) | ReGen3D · Voxel held out, reconstruction | CD (x10^-3): 50.6 -> 33.1 | Improved |
| add ERA on top of URT and SMI (ablation) | ReGen3D · Voxel held out, segmentation | mIoU (%): 58.6 -> 76.6 | Improved |
| add ERA on top of URT and SMI (ablation) | ReGen3D · Training representations (P, M, G) | reconstruction/classification/segmentation: 23.4 23.6 29.2 unchanged | No clear effect |
| replace representation-specific models with ReGen3D unified model | ReGen3D vs ShapeSplat · RepShift-3D unseen Gaussian, classification | accuracy (%): 45.2 -> 85.1 | Improved |
| replace naive cross-attention baseline with ReGen3D | ReGen3D vs Baseline · RepShift-3D unseen Mesh reconstruction | CD (x10^-3): 52.4 -> 32.1 | Improved |
| use ERA to adapt to real-scan point clouds | ReGen3D · ScanObjectNN | accuracy (%): 70.6 Acc., 40.1 CD | No clear effect |
Closest papers by meaning
- Uni3D: Exploring Unified 3D Representation at Scale (ICLR 2024 · similarity 0.88)
- Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering (NeurIPS 2026 · similarity 0.88)
- g3D-LF: Generalizable 3D-Language Feature Fields for Embodied Tasks (CVPR 2025 · similarity 0.88)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.88)
- LaGeM: A Large Geometry Model for 3D Representation Learning and Diffusion (ICLR 2025 · similarity 0.88)
- A Unified Framework for 3D Scene Understanding (NeurIPS 2024 · similarity 0.88)
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.
