Sparse depth map reconstruction from image sequences of the buildings
Zheng Fa Hu, Min Sun, Kerui Xia · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
This paper presents an improved cooperative matching algorithm based on feature regions. The proposed algorithm uses normal vector in each pixel position of the whole images to extract the feature regions of image sequences of buildings. In order to obtain a sparse depth map, matching areas are limited to feature regions rather than whole images. The image similarity of each pixel is defined as match value. Iteratively update match values and make the match values convergent. For each pixel, the pixel and disparity which the maximum match value corresponds to are regarded as matching results. By using feature regions extraction of image sequences, not only can the reconstruction process be further simplified, the running speed can also be increased. The experimental results show that our method is effective and can avoid mismatching in some regions which are texture-less or have sparse texture. Meanwhile, the problem of large computation is solved by pruning unnecessary matching regions.