Driv3R: Learning Dense 4D Reconstruction for Autonomous Driving

For real-time 4D reconstruction in autonomous driving, existing DUSt3R-based methods fall short due to inaccurate modeling of fast-moving dynamic objects and reliance on computationally expensive global alignment processes. In comparison, our Driv3R predicts per-frame pointmaps in the global consistent coordinate system in a optimization-free manner, and can accurately reconstruct fast-moving objects on large-scale scenes with 15x faster inference speed compared to methods requiring global alignment.

Dynamic Scenes in NuScenes

Driv3R outperforms previous frameworks in 4D dynamic scene reconstruction, achieving 15x faster inference speed compared to methods requiring global alignment.

Point Cloud

Depth Estimation

Driv3R achieves better representations of fast-moving objects and ensures both temporal and spatial consistency in the global coordinate system without any optimization or alignment.