Research map / Detection and segmentation / 3D detection and driving
Gaussian-Det: Learning Closed-Surface Gaussians for 3D Object Detection
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 Gaussian-Det, which treats 3D Gaussian Splatting primitives as surface descriptors and uses a Closure Inferring Module (flux-based surface-closure prior plus variational feature residuals) to refine multi-view 3D object proposals on 3D-FRONT and ScanNet.
Problem
Multi-view RGB 3D object detection in indoor scenes is ill-posed; existing NeRF-based and projection-based detectors use discrete sampling that fails to capture true objectness, and Gaussian Splatting reconstructions contain many outlier Gaussians.
Approach
Represent each Gaussian as a surface element (center, normal from covariance eigenvector, area), then score proposals by the flux of a constant vector field through the enclosed partial surfaces (near zero for closed objects), using this closure weight to gate holistic proposal features; a VAE-style variational residual handles the underdetermined Gaussian features.
Key results
- On 3D-FRONT, Gaussian-Det reaches AP25 96.7 and AP50 77.7, versus 88.2 / 71.0 for the second-best G-BRNet.
- On ScanNet, AP25 is 71.7 and AP50 is 24.5, versus 66.8 / 20.4 for the best baselines among those listed.
- Gaussian-Det runs at 33.3 FPS with a 24.0MB model, versus 1.51 FPS for NeRF-RPN and 1.11 FPS for NeRF-MAE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Gaussian-Det instead of the strongest listed baseline | 3D-FRONT detection · 3D-FRONT | AP@0.25: 96.7 vs 88.2 | Helped |
| use Gaussian-Det instead of the strongest listed baseline | ScanNet detection · ScanNet | AP@0.25: 71.7 vs 66.8 | Helped |
| use Gaussian-Det instead of the strongest listed baseline | ScanNet detection · ScanNet | AR@0.25: 87.3 vs 92.0 | Hurt |
| add holistic surface coalescence (ablation) | Gaussian-Det · 3D-FRONT | AP@0.25: 92.8 -> 96.7 | Helped |
| add holistic surface coalescence (ablation) | Gaussian-Det · 3D-FRONT | AP@0.50: 73.9 -> 77.7 | Helped |
| use closure-based scoring instead of setting the closure score to 1 (ablation) | Gaussian-Det · 3D-FRONT | AP@0.25: 93.5 -> 96.7 | Helped |
| use closure-based scoring instead of setting the closure score to 1 (ablation) | Gaussian-Det · 3D-FRONT | AP@0.50: 75.0 -> 77.7 | Helped |
| add probabilistic feature residuals (ablation) | Gaussian-Det · 3D-FRONT | AP@0.25: 95.7 -> 96.7 | Helped |
| add probabilistic feature residuals (ablation) | Gaussian-Det · 3D-FRONT | AP@0.50: 74.7 -> 77.7 | Helped |
| learn flexible gamma values rather than using fixed gamma=0.5 (ablation) | Gaussian-Det · 3D-FRONT | AP@0.25: 96.4 -> 96.7 | Helped |
| learn flexible gamma values rather than using fixed gamma=0.5 (ablation) | Gaussian-Det · 3D-FRONT | AP@0.50: 76.2 -> 77.7 | Helped |
| use proposal-center-oriented surface normals instead of the alternative orientation strategy (ablation) | Gaussian-Det · 3D-FRONT | AP@0.50: 75.7 -> 77.7 | Helped |
Surprising findings
- Plugging RepSurf-U into G-BRNet lowers AP25 on ScanNet from 63.1 to 47.7, suggesting local surface operators are distracted by Gaussian outliers.
- Point-cloud detectors applied directly to Gaussians already reach performance comparable to multi-view and NeRF-based methods.
Limitations
- Struggles with heavily occluded objects, which is inherent to multi-view detection.
- Jointed objects can form a closed surface and be detected as a single object.
- Closure prior weakens for objects with low exposure to the camera, such as distant or partly hidden ones.
Experimental setup
Datasets: 3D-FRONT ScanNet
Benchmarks: 3D-FRONT ScanNet Gaussian-Grouping open-vocabulary 3D instance segmentation
Metrics: AP@0.25 AP@0.5 AR@0.25 AR@0.5 mIoU mBIoU FPS model size
Baselines: ImVoxelNet NeRF-RPN NeRF-MAE G-VoteNet G-GroupFree G-FCAF3D G-BRNet RepSurf-U
Builds on: 3D Gaussian Splatting VoteNet NeRF-RPN PointNet++ SuGaR
Closest papers by meaning
- GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)
- Mitigating Ambiguities in 3D Classification with Gaussian Splatting (CVPR 2025 · similarity 0.86)
- HUGS: Holistic Urban 3D Scene Understanding via Gaussian Splatting (CVPR 2024 · similarity 0.85)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.85)
- Rethinking End-to-End 2D to 3D Scene Segmentation in Gaussian Splatting (CVPR 2025 · similarity 0.85)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.85)
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.