Research map / 3D vision / Gaussian splatting
DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving
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 DrivingRecon, a feed-forward large reconstruction model that predicts 4D Gaussians from surround-view driving videos, using a Prune and Dilate Block to remove redundant cross-view Gaussians and dynamic/static decoupled supervision.
Problem
Photorealistic 4D reconstruction of street scenes for driving simulation is slow because existing methods require per-scene iterative optimization, and feed-forward reconstruction models do not handle surround-view redundancy, sparse supervision, and dynamic objects.
Approach
A feed-forward 4D Gaussian model for surround-view video with a PD-Block that prunes overlapping cross-view and background Gaussians while dilating points for complex objects, plus static/dynamic rendering supervision across timestamps using predicted optical flow.
Key results
- On Waymo NOTA-DS64 reconstruction, DrivingRecon reports PSNR 23.70 and LPIPS 0.17 versus 21.33 and 0.24 for the best baseline MVSplat.
- On the generalization set (Diversity-54), novel view PSNR reaches 21.41 versus 19.63 for MVSplat.
- Ablation removing the PD-Block drops reconstruction PSNR from 22.73 to 19.27 on Diversity-54.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DrivingRecon instead of MVSplat | DrivingRecon · Waymo NOTA-DS64 reconstruction | PSNR: 21.33 -> 23.70 | Helped |
| Use DrivingRecon instead of MVSplat | DrivingRecon · Waymo NOTA-DS64 novel view synthesis | LPIPS: 0.28 -> 0.21 | Helped |
| Use DrivingRecon instead of MVSplat | DrivingRecon · Diversity-54 reconstruction | PSNR: 20.53 -> 22.73 | Helped |
| Use DrivingRecon instead of MVSplat | DrivingRecon · Diversity-54 novel view synthesis | PSNR: 19.63 -> 21.41 | Helped |
| Add PD-Block (ablation) | DrivingRecon · Diversity-54 reconstruction | PSNR: 19.27 -> 22.73 | Helped |
| Add DS-R (ablation) | DrivingRecon · Diversity-54 reconstruction | PSNR: 21.44 -> 22.73 | Helped |
| Add 3D-PE (ablation) | DrivingRecon · Diversity-54 reconstruction | PSNR: 21.65 -> 22.73 | Helped |
| Add TCA (ablation) | DrivingRecon · Diversity-54 reconstruction | PSNR: 20.10 -> 22.73 | Helped |
| Increase training samples from 32 to 512 | DrivingRecon · Diversity-54 reconstruction | PSNR: 21.47 -> 24.25 | Helped |
| Increase segmentation loss weight (ablation) | DrivingRecon · Reconstruction | reconstruction performance | Hurt |
| Add depth supervision (ablation) | DrivingRecon · Reconstruction | reconstruction quality | Helped |
Limitations
- Relies on LiDAR-derived depth supervision and segmentation labels from DeepLabv3+ and SAM for training.
- Evaluation is on a small set of Waymo scenes (64 and 54 scenes), and generalization is tested on the same dataset family.
- Some baseline comparisons and downstream results rely on fine-tuning setups with limited reported details.
Experimental setup
Datasets: Waymo Open dataset (NOTR subset NOTA-DS64) Diverse-56 / Diversity-54 nuScenes
Benchmarks: Waymo NOTA-DS64 Diversity-54 (Diverse-56) nuScenes UniAD/ViDAR downstream tasks
Metrics: PSNR SSIM LPIPS mAP NDS AMOTA minADE avg.L2
Baselines: LGM pixelSplat MVSplat L4GM DG-BEV PD-BEV UniAD ViDAR
Builds on: LGM pixelSplat MVSplat L4GM 3D Gaussian Splatting
Closest papers by meaning
- DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input (AAAI 2025 · similarity 0.89)
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.89)
- DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.87)
- One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving Reconstruction (NeurIPS 2026 · similarity 0.86)
- FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views (NeurIPS 2026 · similarity 0.86)
- GaussianFusion: Gaussian-Based Multi-Sensor Fusion for End-to-End Autonomous Driving (NeurIPS 2025 · 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.