Research map / 3D vision / 3D generation
UniDrive: Towards Universal Driving Perception Across Camera Configurations
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 UniDrive, which warps multi-camera images into a set of unified virtual cameras via ground-aware projection, with CMA-ES optimization of virtual camera configurations, to let BEV 3D detectors generalize across camera configurations on a new CARLA dataset.
Problem
Vision-centric 3D driving perception degrades sharply when camera intrinsics or extrinsics differ from the training configuration, so models do not transfer across vehicle platforms.
Approach
Re-project all original camera views into a unified set of virtual cameras using a ground-aware depth assumption (ground plane near the vehicle, cylindrical surface beyond a distance threshold), and choose the virtual camera parameters by CMA-ES to minimize a distance-weighted angular projection error summed across camera systems; the module is plug-and-play before a 3D detector.
Key results
- UniDrive trained on 5x70+110 reaches 68.8 mAP on the same configuration, versus 63.9 for BEVFusion-C.
- BEVFusion-C trained on 6x80a and tested on 6x60 gets only 1.8 mAP, while UniDrive degrades by at most 9.8% under the largest intrinsic differences.
- Under camera height variation, UniDrive drops only 3.0% while BEVFusion-C drops more than 10%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add UniDrive virtual camera projection to BEVFusion-C | BEVFusion-C · CARLA 3D detection, train 5×70° + 110°, test 5×70° + 110° | mAP: 63.9 -> 68.8 | Helped |
| Add UniDrive virtual camera projection | BEVFusion-C · CARLA 3D detection; train 6 × 60°, test 5×70° + 110° | mAP: 4.6 -> 63.6 | Helped |
| Add UniDrive virtual camera projection | BEVFusion-C · CARLA 3D detection; train 6 × 70°, test 6 × 60° | mAP: 9.2 -> 63.1 | Helped |
| Optimize the virtual camera configuration rather than using the intuitive unoptimized configuration (ablation) | UniDrive · CARLA 3D detection across camera configurations | mAP | Helped |
| Use an intuitive virtual camera configuration without optimization (ablation) | UniDrive · CARLA 3D detection across camera configurations | mAP | Mixed |
| Add UniDrive for varying camera heights | BEVFusion-C · CARLA 3D detection across camera heights | performance: more than 10% -> 3.0% | Helped |
| Add UniDrive for camera placement changes | BEVFusion-C · CARLA 3D detection; train 6 × 80° b, test 6 × 80° a | performance drop: 5.9% -> 4.6% | Helped |
Limitations
- Only a limited set of camera configurations was tested and may not cover all real-world setups.
- All experiments are on simulated CARLA data; no real-world validation.
- Ground-aware assumption relies on a flat ground plane and distance threshold, and some regions are unwarped where cameras lack coverage.
Experimental setup
Datasets: CARLA-generated multi-camera dataset (nuScenes format)
Benchmarks: CARLA multi-camera configuration dataset
Metrics: mAP class-level AP
Baselines: BEVFusion-C
Builds on: BEVFusion CMA-ES nuScenes
Closest papers by meaning
- VR-Drive: Viewpoint-Robust End-to-End Driving with Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- VGGDrive: Empowering Vision-Language Models with Cross-View Geometric Grounding for Autonomous Driving (CVPR 2026 · similarity 0.86)
- UniPAD: A Universal Pre-training Paradigm for Autonomous Driving (CVPR 2024 · similarity 0.86)
- UniScene: Unified Occupancy-centric Driving Scene Generation (CVPR 2025 · similarity 0.86)
- ReSimAD: Zero-Shot 3D Domain Transfer for Autonomous Driving with Source Reconstruction and Target Simulation (ICLR 2024 · similarity 0.85)
- Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene (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.