PGA-Stereo: Robust Stereo Matching by Fusing Monocular Semantic Priors with Mixture-of-Experts Aggregation

The core bottlenecks hindering the real-world deployment of stereo matching are the inherent matching ambiguity in ill-posed regions, such as reflective and texture-less surfaces, and the lack of robustness under physical degradation scenarios, like sensor failures. While recent methods leverage semantic priors from monocular Vision Foundation Models (VFMs), their inherent deficiency in geometric constraints, coupled with the fragility of a unimodal architecture, fails to fundamentally address these challenges. We therefore propose PGA-Stereo, a novel unified framework for stereo matching that, for the first time, synergizes monocular and binocular depth estimation within a single architecture. In our framework, a Prior-Guided Feature Modulator (PGFM) leverages monocular semantic priors as conditional guidance to enhance the discriminability of binocular geometric features at the feature fusion stage. To improve the adaptability of cost aggregation, we devise an Adaptive Expert Cost Aggregation Module (AECAM) that performs scene-aware, expert-based processing for ill-posed regions characterized by distinct information biases. Moreover, to counteract the potential loss of information in critical regions during iterative optimization, a Dynamic Fusion Optimization Module (DFOM) adaptively integrates multi-source expert information within a ConvGRU-based iterative update scheme, enabling efficient disparity refinement. Extensive experiments demonstrate that PGA-Stereo is highly effective for ill-posed scenarios and exhibits superior generalization capabilities in real-world settings. At the time of submission, PGA-Stereo was recognized as the leading method among all published approaches on the KITTI 2012 Reflective benchmark. PGA-Stereo establishes a new state-of-the-art (SOTA) on the KITTI 2015 benchmark, achieving a 3-pixel outlier rate of 2.25% in foreground regions (D1-fg) and outperforming IGEV-Stereo by 15.73%.