CVPR 2026 Paper Spotlight: DriveMVS — LiDAR-Prompted Spatio-Temporal Multi-View Stereo for Autonomous Driving

CVPR 2026 — Sun Qihao and colleagues from Harbin Institute of Technology and collaborating institutions published the paper “LiDAR Prompted Spatio-Temporal Multi-View Stereo for Autonomous Driving” (DriveMVS), proposing a novel framework that deeply integrates sparse LiDAR with multi-view stereo vision, achieving breakthrough results in depth estimation for autonomous driving.

The Core Problem

Perception and simulation for autonomous driving heavily depend on accurate metric depth information, but existing approaches face three interrelated challenges:

1. Insufficient Metric Accuracy: Pure vision methods lack absolute scale anchoring
2. Poor Consistency: Depth estimates are inconsistent across views and over time
3. Weak Cross-Domain Generalization: Performance degrades when transferring from training scenarios to new environments

DriveMVS Technical Approach

DriveMVS resolves these contradictions through two core insights:

1. LiDAR as Geometric Prompt

Sparse but metrically accurate LiDAR point clouds are embedded into an MVS cost volume, providing guidance in two complementary ways:

• Hard Geometric Prior: Directly anchors absolute scale during cost volume construction
• Soft Feature Guidance: Deeply integrates LiDAR features with visual features through a Triple-Cue Combiner

2. Spatio-Temporal Decoder

Unlike traditional per-frame independent depth estimation, DriveMVS employs a joint spatio-temporal decoder that simultaneously leverages geometric cues from the MVS cost volume and temporal context from adjacent frames, ensuring depth estimation consistency and stability across the temporal dimension.

Experimental Results

DriveMVS has achieved state-of-the-art performance on multiple public benchmarks:

• Metric Accuracy: Surpasses all existing methods across every evaluation metric
• Temporal Stability: Exceptionally strong inter-frame depth consistency, significantly reducing the “flickering” phenomenon
• Zero-Shot Cross-Domain Transfer: Maintains high performance on unseen datasets without any fine-tuning

Significance for the Multi-Camera Vision Industry

DriveMVS’s core contribution lies in demonstrating that sparse LiDAR + dense multi-view stereo vision represents a highly complementary fusion pathway: LiDAR provides absolute scale “anchors,” multi-view vision delivers dense spatial coverage, and the spatio-temporal decoder eliminates inter-frame jitter. This paradigm carries important implications for the broader multi-camera industry:

1. Sensor fusion is not about simple stacking, but deep coupling: DriveMVS performs fusion at the feature level rather than the output level, yielding vastly superior results compared to post-processing concatenation
2. Spatio-temporal consistency is critical for industrial deployment: Applications such as autonomous driving and robot navigation cannot tolerate inter-frame depth jumps; spatio-temporal decoder thinking can be applied to backend optimization across various multi-camera vision systems
3. “Sparse + Dense” combinations maximize cost-effectiveness: Rather than upgrading to high-beam-count LiDAR across the entire stack, low-beam-count LiDAR + high-quality multi-camera vision can achieve comparable results at a fraction of the cost

The paper code is open-sourced: github.com/Akina2001/DriveMVS