Construction of video mosaics using the Minimum Description Length

The paper presents work on the construction of video mosaics of the sea bottom, acquired by an autonomous underwater robot. The approach presented, based on the Minimum Description Length, while strongly grounded on probabilistic theory, is able to solve the image registration problem without need for user-defined probabilistic models (no need for neither world nor sensor models), presenting several advantages over traditional techniques. Mosaic construction is based on a dense locally estimated pure translational displacement field. The fact that the proposed metric for image co-registration does not require the presence of distinctive features (like corner-like structures), together with the fact that it is inherently robust to smooth illumination variations across the image plane, makes it particularly suitable for the underwater environment, where individual features present most often a strong resemblance, and where illumination variations are known to be frequent, due to the need to carry the light source onboard the platform. Even if a pure translational model is locally fit to each small image region, the method can globally cope with strong orientation variations of the platform, as the results present demonstrate. The paper details the formal presentation of the criterion used for local displacement estimation, and shows results of mosaic construction with real underwater images acquired under challenging conditions