Research map / 3D vision / Gaussian splatting
Deformable Radial Kernel 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 Deformable Radial Kernel (DRK), a generalized planar splatting primitive with learnable radial bases, L1/L2 blended distance, and edge sharpening, that represents complex shapes with far fewer primitives than Gaussian splatting.
Problem
Gaussian splatting primitives are radially symmetric and smooth, so representing sharp edges and diverse shapes requires thousands of Gaussians, causing inefficiency in memory and computation.
Approach
Replace the Gaussian kernel with a planar deformable radial kernel defined by learnable radial basis lengths and polar angles, blending L1 and L2 norms with a learnable weight, and applying a piecewise linear sharpening function; includes ray-primitive intersection depth sorting, polygon-based tile culling, and cache sorting.
Key results
- On DiverseScenes, DRK achieves average PSNR 37.58 and LPIPS 0.0564 versus 35.68 PSNR for 3D-HGS and 34.41 for 3D-GS.
- DRK (S2) uses 42K primitives on DiverseScenes versus roughly 336K-373K for the Gaussian baselines, with 12.3 MB model size.
- On Mip-NeRF360, DRK reaches LPIPS 0.2364 and SSIM 0.7871 versus 0.3045 and 0.7537 for 3D-GS, but PSNR 26.76 is only slightly above 3D-GS (26.48) with more primitives (952K vs 811K).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Gaussian-based splatting with DRK | 3D scene rendering · DiverseScene average | PSNR: 35.68 -> 37.58 | Helped |
| replace Gaussian-based splatting with DRK | 3D scene rendering · DiverseScene average | primitive count: 373K -> 260K | Mixed |
| use the sparser S1 density setting (ablation) | DRK · DiverseScene average | PSNR: 37.58 -> 36.62 | Hurt |
| use the sparser S2 density setting (ablation) | DRK · DiverseScene average | PSNR: 37.58 -> 35.03 | Hurt |
| use the sparser S1 density setting (ablation) | DRK · DiverseScene | primitive count: 260K -> 109K | Helped |
| use the sparser S2 density setting (ablation) | DRK · DiverseScene | primitive count: 260K -> 42 K | Helped |
| replace Gaussian-based splatting with DRK | 3D scene rendering · Mip-NeRF360 | LPIPS: .2688 -> .2364 | Helped |
| replace Gaussian-based splatting with DRK | 3D scene rendering · Mip-NeRF360 | PSNR: 26.94 -> 26.76 | Hurt |
| use the sparser S2 density setting (ablation) | DRK · Mip-NeRF360 | PSNR: 26.76 -> 26.20 | Hurt |
| use the sparser S1 density setting (ablation) | DRK · Mip-NeRF360 | SSIM: .7871 -> .7722 | Hurt |
Surprising findings
- On Mip-NeRF360 the full DRK model has lower PSNR than some baselines on the full image and the authors attribute this to overfitting distant regions, motivating a masked M-PSNR metric.
Limitations
- Full DRK model has larger primitive count and model size than Gaussian baselines and lower FPS (77.5 vs 247 for 3D-GS).
- On unbounded Mip-NeRF360 scenes, PSNR advantage is small or absent, especially for sparser kernel variants, with suspected overfitting in distant regions.
Experimental setup
Datasets: DiverseScenes Mip-NeRF360
Benchmarks: DiverseScenes Mip-NeRF360
Metrics: PSNR SSIM LPIPS M-PSNR primitive count model size FPS
Baselines: 3D-GS 2D-GS 3D-HGS GES 3D-GS (L)
Builds on: 3D Gaussian Splatting 2D Gaussian Splatting EWA splatting 3D-HGS GES
Closest papers by meaning
- 3D Skew Normal Splatting (NeurIPS 2026 · similarity 0.89)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · similarity 0.89)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.88)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Hardware-Rasterized Ray-Based Gaussian Splatting (CVPR 2025 · similarity 0.88)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · 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.