Object Recognition Using Flexible Groups of Local Features

We propose a new object recognition system based on local features and their flexible spatial configuration within a probabilistic framework. Although the training set consists of only 1 image of the object, the system is able to recognize that same object under large rigid and/or non-rigid deformations. This is possible due to the robustness of the local features and of the pairwise geometric constraints between them that we introduce in this paper. These pairwise constraints are also used for grouping features during hypothesis generation. The new grouping method generally produces fewer groups, where each group has more inliers, than the commonly used alternative method. Furthermore, we also propose a novel filtering procedure to select the local features that have high distinctiveness, detectability and robustness to image deformations. As a result, we decrease the ambiguity of matching and, consequently, improve the scalability of the recognition process with respect to the size of the database. We show the viability of this system using a database of 15 objects, and several challenging test images containing rigid/non-rigid deformations, illumination changes, clutter, and partial occlusion. Moreover, we also show the system working in a long range motion problem using a challenging sequence of images.