Unnecessary Image Pair Detection for a Large Scale Reconstruction

This paper proposes an algorithm to detect unnecessary image pairs for efficient structure from motion. Since image pair with small baseline is considered as a poor condition for reconstruction, we focus on computing cameras closely located. We address a term, “remoteness” which indicates the distance between two images in this paper. The remoteness is not affected by image’s intrinsic parameters because camera intrinsic matrix is applied to put the extracted features in the normalized coordinate. The remoteness is computed using feature disparity in normalized coordinate. Therefore, we can detect redundant image pair captured at the near position without reconstruction. The proposed algorithm is proved by experimental results with Notre Dame images.