LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching

A new stereo matching method called LinStereo uses a linear-complexity global attention mechanism to improve depth estimation, particularly in challenging underwater environments, according to research submitted in 2026 [1]. The method, built upon the Depth Anything V3 vision foundation model, introduces a Position-Aware Linear Attention (PALA) module that replaces local recurrence with global aggregation at linear cost [1]. This allows the system to propagate reliable estimates from well-matched regions into degraded areas while preserving disparity structure [1]. The researchers identify that existing iterative stereo pipelines under-exploit three information pathways: multi-scale backbone features are collapsed into single-level correlations, geometric priors remain untapped at initialization, and context propagates only locally [1]. These gaps widen under degraded photometric cues, making underwater scenes a stringent generalization test [1]. To address this, PALA is made effective by two enabling components: Hierarchical Semantic Cost Volumes (HSCV), which supply scale-aligned correlations from the VFM feature hierarchy, and a Depth Prior Initialization (DPI) that converts monocular depth into a metrically calibrated warm start [1]. LinStereo achieves state-of-the-art-level accuracy on standard benchmarks and strong cross-domain generalization [1]. On the TartanAir-UW dataset, the method attained a 28% lower Absolute Relative error (AbsRel), and on SQUID, a real-world underwater dataset, it achieved a 26% improvement [1]. The research was submitted to arXiv on June 24, 2026 [1].