EXPERIMENT ON PRODUCING DISPARITY MAPS FROM AERIAL STEREO IMAGES USING UNSUPERVISED AND SUPERVISED METHODS

Abstract

Recent advancement in hardware and software provides the possibility of realizing full automation in stereo-image tasks. This paper investigated disparity map generation from aerial images with different methods: unsupervised method and supervised methods. The datasets were from aerial stereo dense matching benchmark dataset for deep learning in ISPRS 2021: Vaihingen dataset and the WHU MVS/Stereo Dataset released in the CVPR 2020. Two neural networks: GC-net and PSMnet have been trained with the Vaihingen dataset and the WHU MVS/Stereo Dataset. With unsupervised methods, stereo block matching(StereoBM) and Stereo Semi-Global Matching (StereoSGM) methods from the OpenCV were studied. We selected seven image pairs from the Vaihingen dataset and six image pairs from the WHU dataset for testing and evaluation. Difficulty scenes such as textureless areas, reflective surfaces, and repetitive patterns were also included in our study. The performance from different methods was compared by both visualization and quantitative means. The advantages and disadvantages are presented.