Research map / 3D vision / 3D generation
Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding
Kadir Yilmaz, Adrian Kruse, Tristan Höfer, Daan de Geus, Bastain Leibe
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 data-efficient recipe (augmentation, regularization, CNN distillation) (ablation) | Volt · ScanNet200 val | mIoU: 31.0 -> 40.0 | Improved |
| add the same recipe upgrades to PTv3 (ablation) | PTv3 · ScanNet200 val | mIoU: 35.2 -> 37.5 | Improved |
| add joint multi-dataset training | Volt-S · ScanNet val | mIoU: 77.2 -> 80.2 | Improved |
| scale backbone from Volt-S to Volt-B under joint training | Volt · ScanNet200 val | mIoU: 38.1 -> 40.0 | Improved |
| use asymmetric instead of symmetric RoPE allocation (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 35.4 -> 36.2; 69.5 -> 70.5 | Improved |
| keep metric token indices instead of per-scene normalization (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 35.7 -> 36.2; 69.3 -> 70.5 | Improved |
| use RoPE instead of Fourier positional encoding (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 28.6 -> 36.2; 66.0 -> 70.5 | Improved |
| add lightweight decoder (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 35.0 -> 36.2; 69.9 -> 70.5 | Improved |
| use larger decoder with two residual conv blocks (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 36.2 -> 35.8; 70.5 -> 69.5 | Hurt |
| use smaller 3x3x3 patches instead of 5x5x5 (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 36.2 -> 37.7; 70.5 -> 71.0 (27 ms -> 67 ms) | Mixed |
| use larger 7x7x7 patches instead of 5x5x5 (ablation) | Volt-S · ScanNet200 / SemanticKITTI | mIoU: 36.2 -> 33.5; 70.5 -> 69.6 | Hurt |
| replace MinkUNet backbone with Volt-S in SPFormer | SPFormer · ScanNet200 val | mAP50: 33.8 -> 48.4 | Improved |
Closest papers by meaning
- Multi-View 3D Reconstruction With Transformers (ICCV 2021 · similarity 0.89)
- Video-3D LLM: Learning Position-Aware Video Representation for 3D Scene Understanding (CVPR 2025 · similarity 0.87)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- H2GFormer: Horizontal-to-Global Voxel Transformer for 3D Semantic Scene Completion (AAAI 2024 · similarity 0.87)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.86)
- Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning (CVPR 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.
