Research map / 3D vision / 3D generation
3DMV: Joint 3D-Multi-View Prediction for 3D Semantic Scene Segmentation
Angela Dai, Matthias Niessner
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 RGB 2D-feature input to the 3D geometry network (1 view) (ablation) | 3DMV · ScanNet | average accuracy: 54.4 -> 70.1 | Improved |
| use per-voxel colors instead of a learned 2D feature extractor (ablation) | 3D geometry-only network · ScanNet | average accuracy: 54.4 -> 55.9 | No clear effect |
| add 3D geometry input to backprojected 2D features (3 views) (ablation) | 3DMV · ScanNet | average accuracy: 58.2 -> 73.0 | Improved |
| add 3D convolutions after backprojection of 2D labels (3 views) (ablation) | 2D label backprojection · ScanNet | average accuracy: 44.2 -> 58.2 | Improved |
| increase views from 0 to 1 (ablation) | 3DMV · ScanNet | average accuracy: 54.4 -> 70.1 | Improved |
| increase views from 1 to 3 (ablation) | 3DMV · ScanNet | average accuracy: 70.1 -> 73.1 | Improved |
| increase views from 3 to 5 (ablation) | 3DMV · ScanNet | average accuracy: 73.1 -> 75.0 | Improved |
| train end-to-end instead of fixed pretrained 2D network (ablation) | 3DMV (1 view) · ScanNet | average accuracy: 69.1 -> 70.1 | Improved |
| train end-to-end instead of fixed pretrained 2D network (ablation) | 3DMV (3 views) · ScanNet | average accuracy: 72.8 -> 73.0 | Improved |
| train end-to-end instead of fixed pretrained 2D network (ablation) | 3DMV (5 views) · ScanNet | average accuracy: 74.5 -> 75.0 | Improved |
| add 2D proxy loss (ablation) | 3DMV · ScanNet | average accuracy | Improved |
| fuse 2D and 3D features at 2/3 of network instead of beginning (ablation) | 3DMV (1 view, fixed 2D) · ScanNet | average accuracy: 67.6 -> 69.1 | Improved |
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.
- ScanComplete — ScanComplete: Large-Scale Scene Completion and Semantic Segmentation for 3D Scans (CVPR 2018) · 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
- 3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans (CVPR 2019 · similarity 0.91)
- Virtual Multi-view Fusion for 3D Semantic Segmentation (ECCV 2020 · similarity 0.91)
- ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes (CVPR 2017 · similarity 0.90)
- SSR-2D: Semantic 3D Scene Reconstruction From 2D Images (TPAMI 2024 · similarity 0.90)
- Depth-aware CNN for RGB-D Segmentation (ECCV 2018 · similarity 0.89)
- DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation (ICLR 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.
