Research map / 3D vision / Gaussian splatting
ProtoCar: Learning 3D Vehicle Prototypes from Single-View and Unconstrained Driving Scene Images
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.
Learns explicit 3D Gaussian vehicle prototypes from single, occluded, unconstrained Waymo driving images using Fibonacci-sphere initialization, tri-plane features and projection-aware feature consistency.
Problem
Single-view 3D vehicle reconstruction from unconstrained real-world driving images with occlusion, varied lighting and viewpoints, where no multi-view supervision is available and existing methods rely on clean multi-view or synthetic data.
Approach
Generates an initial Gaussian point cloud from Fibonacci sphere sampling with two degrees of freedom (scale k and azimuth offset), aligns tri-plane 3D features with visible-region 2D image features via projection-aware consistency, and uses a projection-based mapping mask plus mirror-symmetric camera poses as self-supervision to complete unseen geometry.
Key results
- Initial point cloud CD_raw 0.113 and CD_com 0.095 versus TGS 0.567 and 0.743, with 0.046 s inference versus 0.451 s for TGS.
- On visible-pixel texture reconstruction, PSNR 24.14 and SSIM 0.9058 are comparable to TGS (23.87, 0.9162) and DreamGaussian (25.62, 0.9041); the authors note this comparison is not fair to their method.
- Qualitatively, the method produces plausible full vehicle shapes where Point-E, Shape-E, DreamGaussian and TGS fail under occlusion or poor viewpoints.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProtoCar instead of the strongest listed baseline, DGS | Single-image geometry reconstruction · Single-image geometry reconstruction (raw LiDAR) | CDraw: 0.476 -> 0.113 | Helped |
| Use ProtoCar instead of the strongest listed baseline, Point-E | Single-image geometry reconstruction · Single-image geometry reconstruction (completed LiDAR) | CDcom: 0.672 -> 0.095 | Helped |
| Use ProtoCar instead of the fastest listed baseline, TGS | Single-image geometry reconstruction · Single-image geometry reconstruction | inference time (s): 0.451 -> 0.046 | Helped |
| Use ProtoCar instead of the strongest listed baseline, DGS | Visible-pixel texture reconstruction · Visible-pixel texture reconstruction | PSNR: 25.62 -> 24.14 | Hurt |
| Use ProtoCar instead of the strongest listed baseline, TGS | Visible-pixel texture reconstruction · Visible-pixel texture reconstruction | SSIM: 0.9162 -> 0.9058 | Hurt |
| Use ProtoCar instead of the strongest listed baseline, TGS | Visible-pixel texture reconstruction · Visible-pixel texture reconstruction | LPIPS: 0.2105 -> 0.2141 | Hurt |
Surprising findings
- Restricting the feature and texture losses to visible pixels yields reasonable results despite very low-quality, occluded input data.
- Global pooling of image features is used deliberately to avoid overfitting to visible views and to aid completion of unseen regions.
Limitations
- Ground-truth completed point clouds are noisy (e.g., irremovable ground points) and are generated by a pre-trained completion model.
- Texture metrics are computed only on visible pixels, so they do not fully evaluate full-model quality, and the comparison is acknowledged as unfair.
Experimental setup
Datasets: Waymo Open Dataset Waymo Perception Object Assets subset
Benchmarks: Waymo Perception Object Assets vehicle subset
Metrics: Chamfer distance (raw and completed) inference time PSNR SSIM LPIPS
Baselines: Point-E Shape-E DreamGaussian TGS AutoRF
Builds on: 3D Gaussian splatting TGS triplane representation Chamfer distance point cloud supervision
Closest papers by meaning
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.88)
- UrbanCAD: Towards Highly Controllable and Photorealistic 3D Vehicles for Urban Scene Simulation (CVPR 2025 · similarity 0.88)
- DrivingSphere: Building a High-fidelity 4D World for Closed-loop Simulation (CVPR 2025 · similarity 0.87)
- Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene (CVPR 2025 · similarity 0.86)
- VR-Drive: Viewpoint-Robust End-to-End Driving with Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.86)
- UniPAD: A Universal Pre-training Paradigm for Autonomous Driving (CVPR 2024 · 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.