Research map / 3D vision / Gaussian splatting
3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes
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 3D Convex Splatting, which uses differentiable 3D smooth convex shapes (convex hulls of point sets with LogSumExp smoothness and sigmoid sharpness) as radiance-field primitives rendered with a CUDA tile rasterizer, outperforming 3DGS on Mip-NeRF360, Tanks and Temples and Deep Blending with fewer primitives.
Problem
Gaussian primitives in radiance field reconstruction struggle to capture hard edges and flat surfaces without many primitives, leading to high memory use and diffuse geometry.
Approach
Replace Gaussian primitives with 3D smooth convexes defined as convex hulls of K freely moving 3D points, projected to 2D convex hulls and rendered via a differentiable tile-based splatting rasterizer, with per-convex smoothness (delta) and sharpness (sigma) parameters and a split-based densification scheme.
Key results
- 3DCS improves PSNR by up to 0.81 and LPIPS by 0.026 over 3DGS while using fewer primitives (abstract).
- On Tanks and Temples 3DCS reaches 23.95 PSNR and 0.157 LPIPS vs 3DGS 23.14 PSNR and 0.183 LPIPS, at 282MB vs 411MB.
- On indoor Mip-NeRF360 scenes 3DCS gets 31.33 PSNR vs 30.41 for 3DGS, with 0.166 vs 0.189 LPIPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 3D Gaussian primitives with smooth convex primitives | 3DCS · Tanks & Temples | PSNR: 23.14 -> 23.95 | Helped |
| replace 3D Gaussian primitives with smooth convex primitives | 3DCS · Deep Blending | PSNR: 29.41 -> 29.81 | Helped |
| replace 3D Gaussian primitives with smooth convex primitives | 3DCS · Mip-NeRF360 | PSNR: 27.21 -> 27.29 | Helped |
| use 3 points per convex instead of the full model's 6 (ablation) | 3DCS · Tanks & Temples | PSNR: 23.90 -> 22.40 | Hurt |
| use 4 points per convex instead of the full model's 6 (ablation) | 3DCS · Tanks & Temples | PSNR: 23.90 -> 23.73 | No clear effect |
| use 5 points per convex instead of the full model's 6 (ablation) | 3DCS · Tanks & Temples | PSNR: 23.90 -> 23.70 | No clear effect |
| use 7 points per convex instead of the full model's 6 (ablation) | 3DCS · Tanks & Temples | PSNR: 23.90 -> 23.90 | No clear effect |
| split each convex into more new convexes (2 or 3 instead of 6) (ablation) | 3DCS · Tanks & Temples | visual quality | Hurt |
Surprising findings
- PSNR favors blurrier images; 3DCS visually closer to ground truth in Flower scene despite lower PSNR in that region.
Limitations
- Slower training and lower rendering FPS than 3DGS
- Outdoor nature scenes show smaller advantage over 3DGS
- Higher memory than Mip-NeRF360 on some datasets
Experimental setup
Datasets: Mip-NeRF360 Tanks and Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks and Temples Deep Blending
Metrics: PSNR SSIM LPIPS training time FPS memory
Baselines: 3DGS GES 2DGS Mip-NeRF360 ZipNeRF Instant-NGP
Builds on: 3D Gaussian Splatting CvxNet Mip-NeRF360 2D Gaussian Splatting
Closest papers by meaning
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.92)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.92)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.91)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
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.