DEEP LEARNING-BASED STEREO MATCHING FOR HIGH-RESOLUTION SATELLITE IMAGES: A COMPARATIVE EVALUATION

Dense matching plays an important role in 3D modeling from satellite images. Its purpose is to establish pixel-by-pixel correspondences between two stereo images. The most well-known algorithm is the semi-global matching (SGM), which can generate high-quality 3D models with high computational efficiency. Due to the complex coverage and imaging condition, SGM cannot cope with these situation well. In recent years, deep learning-based stereo matching has attracted wide attention and shown overwhelming benifits over traditional algorithms in terms of precision and completeness. However, existing models are usually evaluated by using close-ranging datasets. Thus, this study investigates the recent deep learning models and evaluate their performance on both close-ranging and satellite image datasets. The results demonstrate that deep learning network can better adapt to the satellite dataset than the typical SGM. Meanwhile, the generalization ability of deep learning-based models is still low for the real application at recent time.