Research map / Point clouds / Point cloud classification
Mitigating Ambiguities in 3D Classification with Gaussian 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.
Proposes classifying 3D objects from Gaussian Splatting point clouds (position, opacity, scale, rotation) to resolve ambiguities for wire-like vs flat surfaces and transparent/reflective objects, and releases a 4,000-object real-world GS point cloud dataset built from MVImageNet.
Problem
Point cloud-based 3D classification confuses wire-like and flat surfaces and transparent or reflective objects because point clouds lack local geometry (scale/rotation) and soft opacity information.
Approach
Replace the point cloud input with a Gaussian Splatting point cloud whose per-point scale, rotation and opacity encode local surface type and material transparency, keeping classifier architectures unchanged.
Key results
- Adding GS coefficients improves PointNet overall accuracy from 73.56% to 80.87% (+7.31) on the GS point cloud dataset.
- All six classifiers improve with GS input; PTv3 OA rises from 88.78% to 89.93% and PointMLP from 87.91% to 90.21%.
- Opacity specifically helps transparent or reflective objects such as Train and Mug, while scale and rotation help separate Loudspeaker from Can.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GS point cloud coefficients instead of position-only input | PointNet · GS dataset | OA (%): 73.56 -> 80.87 | Helped |
| Use GS point cloud coefficients instead of position-only input | PointNet++ · GS dataset | OA (%): 83.02 -> 86.63 | Helped |
| Use GS point cloud coefficients instead of position-only input | PointNeXt · GS dataset | OA (%): 87.77 -> 89.78 | Helped |
| Use GS point cloud coefficients instead of position-only input | PointMLP · GS dataset | OA (%): 87.91 -> 90.21 | Helped |
| Use GS point cloud coefficients instead of position-only input | DeLA · GS dataset | OA (%): 88.78 -> 90.36 | Helped |
| Use GS point cloud coefficients instead of position-only input | PTv3 · GS dataset | OA (%): 88.78 -> 89.93 | Helped |
| Add opacity to position input (ablation) | PointNet · Train | correct-class probability: 0.69 -> 0.86 | Helped |
| Add scale and rotation to position input (ablation) | PointNet · Loudspeaker | correct-class probability: 0.31 -> 0.60 | Helped |
| Add opacity to position input (ablation) | PointNet · Mug | correct-class probability: 0.46 -> 0.66 | Helped |
Surprising findings
- Gains for state-of-the-art classifiers are smaller than for early methods, with DeLA and PTv3 improving by only about 1-1.6 points OA.
Limitations
- GS point clouds require multi-view RGB captures and per-object GS training (over 600 GPU hours), limiting applicability.
- Only a single 4-channel subset of GS coefficients is used; spherical harmonics are excluded for fair comparison.
- Dataset has only 20 categories and 4,000 objects, derived from one multi-view source.
Experimental setup
Datasets: MVImageNet
Benchmarks: GS point cloud dataset (ours)
Metrics: overall accuracy mean class accuracy
Baselines: PointNet PointNet++ PointNeXt PointMLP DeLA PTv3
Builds on: 3D Gaussian Splatting MVImageNet
Closest papers by meaning
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · similarity 0.89)
- GaussianPainter: Painting Point Cloud into 3D Gaussians with Normal Guidance (AAAI 2025 · similarity 0.88)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.87)
- Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians (CVPR 2025 · similarity 0.87)
- LeanGaussian: Breaking Pixel or Point Cloud Correspondence in Modeling 3D Gaussians (CVPR 2025 · similarity 0.86)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.86)
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.