Unsupervised learning of scene and object planar parts

In this work an adaptive method for accurate and robust grouping of local features belonging to planes of interior scenes and object planar surfaces is presented. For arbitrary set of images acquired from different views, the method organizes a huge number of local SIFT features to fill the gap between low-level vision (front end) and high level vision, i.e. domain specific reasoning about geometric structures. The proposed method consists of three steps: exploration, selection, and merging with verification. The exploration is a data driven technique that proposes a set of hypothesis clusters. To select the final hypotheses a matrix of preferences is introduced. It evaluates each of the hypothesis in terms of number of features, error of transformation, and feature duplications and is applied in quadratic form in the process of maximization. Then, merging process combines the information from multiple views to reduce the redundancy and to enrich the selected representations. The proposed method is an example of unsupervised learning of planar parts of the scene and objects with planar surfaces.