Abstract:
Traditional stereo matching algorithms typically rely on matching of artificially crafted feature descriptors. Nevertheless, their performance in regions like occlusion, weak texture, repetitive texture, and edges leaves much to be desired. In recent years, deep learning-based stereo matching algorithms have emerged as a dominant approach, yielding significant improvements in both performance and robustness. In order to systematically organize the research status and development trends of deep - learning - based stereo matching algorithms, the research status and development trend of stereo matching algorithm based on deep learning, the principle of stereo matching algorithm, traditional stereo matching algorithm, end-to-end and non end-to-end stereo matching algorithm based on deep learning are introduced. Additionally, it comprehensively summarizes relevant training methods, loss functions, and commonly - used public datasets. Meanwhile, a comparative analysis of diverse representative algorithms is conducted on common datasets. Finally, the challenges and prospective development direction of stereo matching algorithm are discussed.