CSStereo: A UAV scenarios stereo matching network enhanced with contrastive learning and feature selection

Abstract

Stereo matching is essential for establishing pixel-level correspondences and estimating depth in scene reconstruction. However, applying stereo matching networks to UAV scenarios presents unique challenges due to varying altitudes, angles, and rapidly changing conditions, unlike the controlled settings in autonomous driving or the uniform scenes in satellite imagery. To address these UAV-specific challenges, we propose the CSStereo network (Contrastive Learning and Feature Selection Stereo Matching Network), which integrates contrastive learning and feature selection modules. The contrastive learning module enhances feature representation by comparing similarities and differences between samples, thereby improving discrimination among features in UAV scenarios. The feature selection module enhances robustness and generalization across different UAV scenarios by selecting relevant and informative features. Extensive experimental evaluations demonstrate the effectiveness of CSStereo in UAV scenarios, and show superior performance in both qualitative and quantitative assessments.