Research map / Detection and segmentation / 3D detection and driving
Any3DIS: Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Any3DIS performs class-agnostic, open-vocabulary and open-ended 3D instance segmentation by tracking 2D masks from SAM 2 across RGB-D frames, lifting them to superpoint-based 3D candidates, and refining them with a dynamic-programming mask optimization.
Problem
Training-free 3D instance segmentation from lifted 2D masks suffers from over-segmentation, redundant 3D proposals, and high runtime due to heuristic unsupervised merging across views.
Approach
Replace dense per-frame 2D mask lifting and heuristic merging with SAM 2 mask tracks seeded from pivot views of sampled superpoints, then select superpoints per track via a view-wise dynamic programming optimization that maximizes inside-mask minus outside-mask point counts.
Key results
- Class-agnostic 3DIS on ScanNet++: AP 22.2 with SAM2-L vs 17.9 for Open3DIS with SAM2-L and 20.7 for Open3DIS with SAM-HQ; 23.1 with SAM2.1-L.
- Runtime 488 s/scene vs 4,226 s/scene for Open3DIS on ScanNet++, with 484 proposals vs 931.
- Open-vocabulary on ScanNet200 (3D + SAM2-L): AP 25.8 vs 23.7 for Open3DIS (3D + SAM-HQ).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Open3DIS with Any3DIS using SAM2-L | Class-agnostic 3DIS · ScanNet++ | AP: 17.9 -> 22.2 | Helped |
| Replace Open3DIS 3D + SAM-HQ proposals with Any3DIS 3D + SAM2-L | Class-agnostic 3DIS · ScanNet200 | AP: 41.5 -> 42.5 | Helped |
| Use Any3DIS instead of Open3DIS | Open-vocabulary 3DIS · ScanNet++ | AP: 11.9 -> 12.9 | Helped |
| Use Any3DIS 3D + SAM2-L proposals instead of Open3DIS 3D + SAM-HQ | Open-vocabulary 3DIS · ScanNet200 | AP: 23.7 -> 25.8 | Helped |
| Use Any3DIS instead of Point-Wise | Open-ended 3DIS · ScanNet++ | AP: 18.4 -> 20.1 | Helped |
| Use Any3DIS instead of Point-Wise | Open-ended 3DIS · ScanNet200 | AP: 16.0 -> 19.1 | Helped |
| Add 3D-Aware 2D Mask Tracking (ablation) | Any3DIS without Mask Optimization · ScanNet++ class-agnostic 3DIS | AP: 17.0 -> 17.9 | Helped |
| Add Mask Optimization (ablation) | Any3DIS baseline without 3D-Aware 2D Mask Tracking · ScanNet++ class-agnostic 3DIS | AP: 17.0 -> 20.3 | Helped |
| Add both 3D-Aware 2D Mask Tracking and Mask Optimization (ablation) | Any3DIS · ScanNet++ class-agnostic 3DIS | AP: 17.0 -> 22.2 | Helped |
| Use top-1 view selection for mask refinement instead of all lifted superpoints (ablation) | Any3DIS Mask Optimization · ScanNet++ class-agnostic 3DIS | AP: 17.9 -> 19.8 | Helped |
| Use top-5 view selection for mask refinement instead of all lifted superpoints (ablation) | Any3DIS Mask Optimization · ScanNet++ class-agnostic 3DIS | AP: 17.9 -> 20.6 | Helped |
| Use top-10 view selection for mask refinement instead of all lifted superpoints (ablation) | Any3DIS Mask Optimization · ScanNet++ class-agnostic 3DIS | AP: 17.9 -> 20.8 | Helped |
Limitations
- The dynamic programming optimization is not guaranteed to find a globally optimal solution because it only adds or skips whole views.
- Performance depends on SAM 2 segmentation and tracking quality, and SAM 2 memory window forgets objects after 7 frames.
Experimental setup
Datasets: ScanNet200 ScanNet++
Benchmarks: ScanNet200 ScanNet++
Metrics: AP AP50 AP25 runtime number of proposals
Baselines: Open3DIS Segment3D SAM-Graph SAM3D SAMPro3D SAI3D OVIR-3D OpenMask3D Search3D MaskClustering
Builds on: Open3DIS SAM 2 SAI3D OVIR-3D Segment3D
Closest papers by meaning
- Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance (CVPR 2024 · similarity 0.91)
- MaskClustering: View Consensus based Mask Graph Clustering for Open-Vocabulary 3D Instance Segmentation (CVPR 2024 · similarity 0.91)
- SAM2Object: Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation (CVPR 2025 · similarity 0.90)
- Segment Any 3D Object with Language (ICLR 2025 · similarity 0.89)
- SAI3D: Segment Any Instance in 3D Scenes (CVPR 2024 · similarity 0.89)
- OnlineAnySeg: Online Zero-Shot 3D Segmentation by Visual Foundation Model Guided 2D Mask Merging (CVPR 2025 · similarity 0.89)
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.