Research map / 3D vision / Gaussian splatting
Fix False Transparency by Noise Guided Splatting
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.
Identifies "false transparency" in object-centric 3D Gaussian Splatting and proposes Noise Guided Splatting (NGS), which fills object interiors with opaque colored noise Gaussians plus an alpha-consistency loss, and a transmittance-based Surface Opacity Score (SOS) to measure it.
Problem
Opaque object surfaces reconstructed with 3D Gaussian Splatting often become falsely semi-transparent, so internal and background Gaussians leak through under camera motion; standard NVS metrics fail to detect this.
Approach
Inject high-opacity noise Gaussians with randomized RGBCMY colors inside the object volume (via convex hull voxelization, depth pruning, multi-scale init, and opacity fine-tuning) to occlude back surfaces, combined with an alpha-consistency loss against object masks, and quantify transparency by rendering the noise infill as a transmittance map.
Key results
- On the Stone dataset, NGS raises SOS from 0.140 (3DGS) to 0.922 while achieving PSNR 34.148 vs 34.610 for 3DGS.
- On DTU, NGS achieves SOS 0.749 vs 0.147 for 3DGS, with PSNR* 25.427 (infill-conditioned) vs 22.967 for 3DGS.
- On OmniObject3D, NGS achieves SOS 0.736 vs 0.215 for 3DGS, with PSNR 33.619 and LPIPS 0.060.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add noise Gaussian infill to GSplat with alpha-consistency loss | GSplat+α · DTU | SOS: 0.598 -> 0.749 | Helped |
| Add noise Gaussian infill to GSplat with alpha-consistency loss | GSplat+α · Stone Dataset | SOS: 0.891 -> 0.922 | Helped |
| Add noise Gaussian infill to GSplat with alpha-consistency loss | GSplat+α · OmniObject3D | SOS: 0.642 -> 0.736 | Helped |
| Add noise Gaussian infill to GSplat with alpha-consistency loss | GSplat+α · DTU | PSNR: 25.435 -> 25.428 | No clear effect |
| Add noise Gaussian infill to GSplat with alpha-consistency loss | GSplat+α · Stone Dataset | PSNR: 33.832 -> 34.148 | Mixed |
| Add noise Gaussian infill to GSplat with alpha-consistency loss | GSplat+α · OmniObject3D | PSNR: 33.575 -> 33.619 | Helped |
| Add binary erosion of noise voxels (ablation) | NGS · Stone Dataset | PSNR: 30.785 -> 32.201 | Helped |
| Add noise Gaussian pruning (ablation) | NGS · Stone Dataset | PSNR: 31.496 -> 32.201 | Helped |
| Add foreground-opacity loss L_f (ablation) | NGS · Stone Dataset | SOS: 0.379 -> 0.969 | Helped |
| Add learning rate reset (ablation) | NGS · Stone Dataset | PSNR: 26.136 -> 32.201 | Helped |
| Use foreground-opacity loss instead of random background (ablation) | NGS · Stone Dataset | SOS: 0.467 -> 0.969 | Helped |
| Randomize noise Gaussian colors during training (ablation) | NGS · Stone Dataset | SOS: 0.962 -> 0.969 | Helped |
Surprising findings
- Adding noise Gaussians improves PSNR and SSIM over the Gsplat+alpha baseline in some cases, suggesting resolving the foreground-background ambiguity also helps convergence.
- Not eroding the noise voxel grid achieves a higher SOS (1.000) but degrades visual quality in the ablation.
Limitations
- Requires high-quality foreground segmentation masks and a reasonable Gaussian point cloud to define the object convex hull.
- Assumes opaque material, so it may wrongly enforce opacity on transparent or translucent objects.
- Increases Gaussian count and memory by roughly 30-50% and is less effective on thin structures and large scenes.
Experimental setup
Datasets: DTU OmniObject3D Stone Dataset
Benchmarks: DTU OmniObject3D Stone Dataset
Metrics: PSNR SSIM LPIPS Surface Opacity Score (SOS) PSNR* SSIM* LPIPS*
Baselines: 3DGS Gaussian Opacity Fields StopThePop Gsplat+alpha
Builds on: 3D Gaussian Splatting Gsplat Gaussian Opacity Fields StopThePop Quickhull
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.89)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.88)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.88)
- Volumetrically Consistent 3D Gaussian Rasterization (CVPR 2025 · similarity 0.88)
- From Chaos to Clarity: 3DGS in the Dark (NeurIPS 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.