Research map / Detection and segmentation / Open-set and few-shot detection
SAMosaic3D: Modular Scene Assembly for Real-Time 3D Segment Anything
Peng Wang, Yongcai Wang, Wang Chen, Hualong Cao, Kang Yang, Chunxu Li, Jie Wen, Deying Li
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 differentiable Fragment-to-Instance module with Lcount (ablation) | SAMosaic3D baseline · ScanNet-25K val | AP: +4.3 AP (40.5 -> 44.8) | Improved |
| add Instance-to-Scene module with two-tier memory and cascaded merging (ablation) | SAMosaic3D · ScanNet-25K val | AP: +3.4 AP (44.8 -> 48.2) | Improved |
| add explicit merge supervision Lmerge (ablation) | SAMosaic3D · ScanNet-25K val | AP: +0.9 AP (48.2 -> 49.1) | Improved |
| replace hard clustering with differentiable soft assignment (ablation) | Fragment-to-Instance · ScanNet-25K val | AP: +2.1 AP (42.5 -> 44.6) | Improved |
| add co-instance affinity matrix W (ablation) | Fragment-to-Instance · ScanNet-25K val | AP: +2.2 AP (44.6 -> 46.8) | Improved |
| add soft-gated attention (ablation) | Fragment-to-Instance · ScanNet-25K val | AP: 46.8 -> 47.3 (uniform weighting) | Improved |
| use max assignment weighting instead of uniform or sum (ablation) | Fragment-to-Instance · ScanNet-25K val | AP: 47.3 / 47.6 -> 49.1 | Improved |
| add two-tier memory over direct geometric merging (ablation) | Instance-to-Scene · ScanNet-25K val | AP: +3.1 AP (43.8 -> 46.9) | Improved |
| cascade semantic matching (S1) with geometric refinement (S2) (ablation) | Instance-to-Scene · ScanNet-25K val | AP: +1.3 AP (46.9 -> 48.2) | Improved |
| reduce short-term memory Nshort from 50 to 20 (ablation) | SAMosaic3D · ScanNet-25K val | AP / latency: 49.1 -> 48.4 AP; 69 -> 58 ms | Mixed |
| increase Nshort from 50 to 100 (ablation) | SAMosaic3D · ScanNet-25K val | AP / latency: 49.1 -> 49.0 AP; 69 -> 88 ms | Hurt |
| use unbounded short-term memory (ablation) | SAMosaic3D · ScanNet-25K val | AP: 49.1 -> 48.0 | Hurt |
Closest papers by meaning
- EmbodiedSAM: Online Segment Any 3D Thing in Real Time (ICLR 2025 · similarity 0.89)
- Online Segment Any 3D Thing as Instance Tracking (NeurIPS 2025 · similarity 0.89)
- Any3DIS: Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking (CVPR 2025 · similarity 0.88)
- SAM2Object: Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation (CVPR 2025 · similarity 0.88)
- OnlineAnySeg: Online Zero-Shot 3D Segmentation by Visual Foundation Model Guided 2D Mask Merging (CVPR 2025 · similarity 0.88)
- Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D Segmentation (CVPR 2025 · similarity 0.87)
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.
