Research map / Point clouds / Point cloud classification
D$^2$GS: Dense Depth Regularization for LiDAR-free Urban Scene Reconstruction
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 D²GS, a LiDAR-free dynamic urban scene reconstruction method that initializes Gaussians from multi-view metric depth with progressive pruning and refines rendered depth jointly with Gaussians via a diffusion-based Depth Enhancer and a road-node planar constraint.
Problem
Urban scene reconstruction with Gaussian splatting usually relies on LiDAR for geometry priors, but accurate LiDAR requires careful calibration and suffers from reprojection errors and missing points, and LiDAR gives only sparse depth supervision.
Approach
Replace LiDAR with dense multi-view metric depth (DepthSplat) back-projected into an initial point cloud pruned progressively by opacity, and jointly refine rendered depth with a Marigold-based diffusion Depth Enhancer (reference, multi-view warping, and smoothness losses) during Gaussian training, plus a road node constrained to be planar.
Key results
- Image reconstruction PSNR 34.34 on Waymo NOTR Dynamic32, versus 33.35 for LiDAR-based OmniRe and 34.07 for OmniRe without LiDAR.
- Depth L1 of 2.6043 versus 5.5345 for OmniRe+DS, about a 53% improvement on depth accuracy as stated in the paper.
- Novel view synthesis PSNR 28.80, versus 29.02 for LiDAR-based OmniRe and 27.69 for OmniRe without LiDAR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use D$^2$GS instead of OmniRe without LiDAR | Waymo NOTR Dynamic32 reconstruction · Waymo NOTR Dynamic32 image reconstruction | PSNR: 34.07 -> 34.34 | Helped |
| Use D$^2$GS instead of OmniRe without LiDAR | Waymo NOTR Dynamic32 reconstruction · Waymo NOTR Dynamic32 novel view synthesis | PSNR: 27.69 -> 28.80 | Helped |
| Use D$^2$GS instead of OmniRe without LiDAR | Waymo NOTR Dynamic32 reconstruction · Waymo NOTR Dynamic32 depth estimation | L1: 4.2197 -> 2.6043 | Helped |
| Add Road Node (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | L1: 2.7580 -> 2.3847 | Helped |
| Add Progressive Pruning (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | L1: 2.9121 -> 2.3847 | Helped |
| Add Depth Enhancer (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | L1: 2.8190 -> 2.3847 | Helped |
| Add warping and smooth losses to Depth Enhancer (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | L1: 2.6477 -> 2.3847 | Helped |
| Use Real-time Update instead of Periodic Update (ablation) | D$^2$GS Depth Enhancer · Waymo NOTR Dynamic32 ablation subset | L1: 2.9168 -> 2.3847 | Helped |
| Use Real-time Update instead of Periodic Update (ablation) | D$^2$GS Depth Enhancer · Waymo NOTR Dynamic32 ablation subset novel view synthesis | PSNR: 26.90 -> 28.04 | Helped |
| Add Road Node (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | PSNR: 32.67 -> 33.63 | Helped |
| Add Progressive Pruning (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | PSNR: 32.25 -> 33.63 | Helped |
| Add Depth Enhancer (ablation) | D$^2$GS · Waymo NOTR Dynamic32 ablation subset | PSNR: 32.77 -> 33.63 | Mixed |
Surprising findings
- Naively supervising OmniRe with DepthSplat depth degrades image reconstruction PSNR (31.26 vs 34.07 without depth), showing noisy metric depth hurts reconstruction.
Limitations
- Assumes accurate camera poses; pose-free operation is future work.
- Depth initialization from MVS is imperfect on dynamic objects and relies on metric depth quality.
- Evaluation limited to the Waymo NOTR Dynamic32 subset; no statistical significance or error bars reported.
Experimental setup
Datasets: Waymo NOTR Dynamic32
Benchmarks: Waymo NOTR Dynamic32
Metrics: PSNR SSIM LPIPS L1 Abs. Rel. RMSE delta < 1.25
Baselines: EmerNeRF 3DGS S3GS PVG OmniRe OmniRe w/o LiDAR OmniRe+DS DepthSplat
Builds on: OmniRe Marigold-DC Marigold DepthSplat
Closest papers by meaning
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.89)
- CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes (ICLR 2025 · similarity 0.89)
- EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · similarity 0.88)
- EAP-GS: Efficient Augmentation of Pointcloud for 3D Gaussian Splatting in Few-shot Scene Reconstruction (CVPR 2025 · similarity 0.88)
- DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.87)
- AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians (NeurIPS 2025 · similarity 0.87)
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.