Research map / 3D vision / 3D generation
GaussianFusion: Gaussian-Based Multi-Sensor Fusion for End-to-End 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 GaussianFusion, which fuses camera and LiDAR features via sparse 2D Gaussians as intermediate carriers and refines trajectories with a cascade planning head, reaching 92.0 PDMS on NAVSIM navtest.
Problem
Multi-sensor fusion for end-to-end autonomous driving: flatten attention fusion lacks interpretability and spatial grounding, while BEV fusion has dense computational and memory overhead.
Approach
Uses sparse, physically parameterized 2D Gaussians as intermediate carriers for camera and LiDAR features, with a dual-branch design (explicit features for scene reconstruction via geometric sampling, implicit features for planning) and a cascade planning head that queries Gaussians to iteratively refine anchor trajectories.
Key results
- Achieves 92.0 PDMS on NAVSIM navtest with V2-99 backbone and 88.8 PDMS with ResNet34, outperforming DiffusionDrive (88.1).
- Reaches 85.0 EPDMS on NAVSIM, outperforming DiffusionDrive (84.0) and ARTEMIS (83.1).
- Achieves 79.1 DS on Bench2Drive, the best among learning-based methods reported, above TF++ (76.9).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GaussianFusion instead of DiffusionDrive with ResNet34 | NAVSIM navtest · NAVSIM navtest | PDMS: 88.1 -> 88.8 | Helped |
| Use GaussianFusion instead of Hydra-MDP++ with V2-99 | NAVSIM navtest · NAVSIM navtest | PDMS: 91.0 -> 92.0 | Helped |
| Use GaussianFusion instead of DiffusionDrive | NAVSIM navtest · NAVSIM navtest | EPDMS: 84.0 -> 85.0 | Helped |
| Use GaussianFusion instead of TF++ | Bench2Drive · Bench2Drive | DS: 76.9±0.9 -> 79.1±1.1 | Helped |
| Add Gaussian explicit fusion (ablation) | TransFuser baseline · NAVSIM navtest | EPDMS: 81.8 -> 84.2 (+2.4) | Helped |
| Add Gaussian implicit fusion to Gaussian explicit fusion (ablation) | TransFuser baseline · NAVSIM navtest | EPDMS: 84.2 -> 84.5 | Helped |
| Add cascade planning to Gaussian explicit fusion (ablation) | TransFuser baseline · NAVSIM navtest | EPDMS: 84.2 -> 84.4 | Helped |
| Combine Gaussian implicit fusion and cascade planning with Gaussian explicit fusion (ablation) | TransFuser baseline · NAVSIM navtest | EPDMS: 84.5 -> 85.0 | Helped |
| Add agent prediction head (ablation) | GaussianFusion · NAVSIM navtest | EPDMS: 85.0 -> 83.8 | Hurt |
| Increase the number of Gaussians from 256 to 512 (ablation) | GaussianFusion · NAVSIM | performance | Helped |
| Increase the number of Gaussians beyond 512 (ablation) | GaussianFusion · NAVSIM | performance | No clear effect |
| Set the number of Gaussian encoder blocks to 4 (ablation) | GaussianFusion · NAVSIM | performance | Helped |
Surprising findings
- Adding the agent prediction head degrades EPDMS (85.0 to 83.8), which the authors attribute to 'Gaussian confusion'.
Limitations
- No quantitative evaluation on tracking small or fast-moving objects.
- Not assessed under sensor occlusion or noise.
- Depends on 3D bounding box and road topology labels for supervision, and customized CUDA operations are not fully optimized.
Experimental setup
Datasets: NAVSIM OpenScene Bench2Drive
Benchmarks: NAVSIM navtest Bench2Drive
Metrics: PDMS EPDMS NC DAC TTC EP LK DS success rate multi-ability mean
Baselines: LTF TransFuser GoalFlow Hydra-MDP Hydra-MDP++ ARTEMIS DiffusionDrive PDM-Lite AD-MLP TCP VAD UniAD
Builds on: TransFuser Hydra-MDP GaussianFormer GaussianAD Deformable attention
Closest papers by meaning
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.87)
- GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction (CVPR 2025 · similarity 0.86)
- DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.86)
- DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving (NeurIPS 2025 · similarity 0.86)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.86)
- GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction (CVPR 2025 · similarity 0.85)
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.