High-speed Motion Analysis with Multi-Exposure Images

We propose a method to estimate dense motion vector fields from multi-exposure images. Our approach relies on finding a sparse set of correspondences between features in a single-exposure image and each exposure in a multi-exposure image using a global optimization technique. We iteratively establish such matches, compute a set of locally restricted transformations for the matches, and construct a dense motion vector field in a multiresolution framework. The estimation of the number of necessary transformations and the regions of influence is guided by superpixel segmentation of the image. We present results for multi-exposure photos of different dynamic scenes.