Research map / Detection and segmentation / CNN features
COB-GS: Clear Object Boundaries in 3DGS Segmentation Based on Boundary-Adaptive Gaussian Splitting
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.
Proposes COB-GS, a 3DGS refinement method that jointly optimizes mask labels and texture with boundary-adaptive Gaussian splitting driven by mask-gradient statistics, yielding clearer object boundaries for 3DGS segmentation.
Problem
3DGS-based object segmentation produces blurry object boundaries because Gaussian primitives have volume and span object edges, and existing methods either ignore semantics during reconstruction or delete boundary Gaussians at the cost of visual quality.
Approach
Uses the sign/consistency of per-view mask-label gradients (a mask-signal statistic, N+ minus N- over N+ plus N-) to identify ambiguous boundary Gaussians and split them along object edges, alternating mask-label optimization with boundary-guided texture restoration, plus removal of tiny boundary Gaussians for robustness to erroneous masks.
Key results
- On NVOS, COB-GS reaches 92.1% mIoU and 98.6% mAcc, versus 91.8% mIoU for FlashSplat and 90.7% for SA3D-GS.
- On NVOS visual quality (CLIP-IQA), COB-GS scores 0.682 / 0.731 / 0.859 across the three boundary and object prompts, the highest among compared methods.
- Ablation on NVOS: boundary-adaptive splitting alone raises mIoU from 91.2 to 91.9; full model with all components reaches 92.1 mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use COB-GS instead of FlashSplat | 3DGS segmentation · NVOS | mIoU (%): 91.8 -> 92.1 | Helped |
| Use COB-GS instead of SA3D-GS | 3DGS segmentation · NVOS | CLIP-IQA: Clear / Unclear Boundary: 0.658 -> 0.682 | Helped |
| Use COB-GS instead of SA3D-GS | 3DGS segmentation · NVOS | CLIP-IQA: Smooth / Noisy Boundary: 0.718 -> 0.731 | Helped |
| Use COB-GS instead of SA3D-GS | 3DGS segmentation · NVOS | CLIP-IQA: Complete / Mutilated Object: 0.835 -> 0.859 | Helped |
| Add boundary-adaptive Gaussian splitting (BAGS) (ablation) | COB-GS · NVOS | mIoU (%): 91.2 -> 91.9 | Helped |
| Add boundary-adaptive Gaussian splitting (BAGS) (ablation) | COB-GS · NVOS | mAcc (%): 98.3 -> 98.5 | Helped |
| Add boundary-guided texture restoration (BGTR) to BAGS (ablation) | COB-GS · NVOS | mIoU (%): 91.9 -> 91.9 | No clear effect |
| Add boundary-guided texture restoration (BGTR) to BAGS (ablation) | COB-GS · NVOS | mAcc (%): 98.5 -> 98.4 | Hurt |
| Add robustness against erroneous masks (RAEM) (ablation) | COB-GS with BAGS and BGTR · NVOS | mIoU (%): 91.9 -> 92.1 | Helped |
| Add robustness against erroneous masks (RAEM) (ablation) | COB-GS with BAGS and BGTR · NVOS | mAcc (%): 98.4 -> 98.6 | Helped |
| Optimize masks only instead of jointly optimizing masks and textures (ablation) | COB-GS · NVOS | PSNR: 23.13 -> 22.61 | Hurt |
| Optimize textures only instead of jointly optimizing masks and textures (ablation) | COB-GS · NVOS | PSNR: 23.07 -> 23.13 | No clear effect |
Surprising findings
- Joint optimization with splitting slightly lowers segmentation accuracy compared with splitting alone, while improving texture quality (PSNR).
- Removing ambiguous Gaussians by simple deletion, as in prior methods, is reported to degrade visual quality.
Limitations
- Floating artifacts from 3D reconstruction are magnified after segmentation and are not addressed.
- Choice of threshold delta affects the number of Gaussians; increasing it greatly increases total Gaussian count.
- Quantitative evaluation is limited to the eight-scene NVOS dataset; visual quality is assessed with a no-reference IQA metric.
Experimental setup
Datasets: NVOS LLFF MIP-360 T&T LERF
Benchmarks: NVOS
Metrics: mIoU mAcc PSNR CLIP-IQA
Baselines: NVOS ISRF SA3D OmniSeg3D SAGD SA3D-GS SAGA FlashSplat
Builds on: 3D Gaussian Splatting SA3D SAM2 Grounding DINO
Closest papers by meaning
- Bootstraping Clustering of Gaussians for View-consistent 3D Scene Understanding (AAAI 2025 · similarity 0.83)
- GaussianCut: Interactive segmentation via graph cut for 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.81)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.81)
- AGILE3D: Attention Guided Interactive Multi-object 3D Segmentation (ICLR 2024 · similarity 0.81)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.80)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.80)
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.