Drive3R: Block-wise Streaming Geometric Foundation Models for Autonomous Driving
Drive3R streams block-wise 3D reconstruction for driving and sets new nuScenes, KITTI, and Waymo results.
Drive3R is a block-wise streaming geometric foundation model for long-horizon multi-view 3D reconstruction in autonomous driving. Block-wise causal attention reduces the cost of fully bidirectional attention while retaining local future context, and confidence-aware re-anchoring resets drifted streaming state before SE(3) alignment. On nuScenes, it improves DVGT-2 from 0.1806 to 0.1081 AbsRel, 0.6750 to 0.5344 point-cloud accuracy, and 1.5812 to 0.9092 ATE. The authors also report state-of-the-art results on KITTI and Waymo.
- Block-wise causal attention avoids full bidirectional attention cost.
- Re-anchoring resets unreliable KV state and aligns segments in SE(3).
- nuScenes AbsRel improves from 0.1806 to 0.1081 versus DVGT-2.
- Authors report state-of-the-art results on nuScenes, KITTI, and Waymo.
From the original article
1 min read at huggingface.co
High-fidelity closed-loop simulation for autonomous driving increasingly relies on dense multi-view 3D reconstruction. While recent feed-forward geometric foundation models have revolutionized per-scene optimization, applying them to continuous, long-horizon driving sequences is computationally intractable due to GPU memory limits and the quadratic complexity of global attention.
By Tianyi Yang, Haojun Yang, Jipeng Xiao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.