Object Recognition by Distortion-Free Graph Embedding and Random Forest

Abstract

Error tolerant graph matching is required not only in many realistic object recognition scenarios, but also in different domains such as document analysis and mechanical drawings. This paper presents such a technique using a distortion-free graph embedding, reformulating the problem as that of finding error-tolerant point matching in the geometric space. The embedding works by finding the distance between every node pair. In order to properly perform the embedding for object recognition in which graph nodes represent image features and graph edges show the relations between the features, we first use a machine learning algorithm, Random Forest, which obtains similar sets of features from images taken from close view points. Given a set of such features, we then perform their consistent ordering based on the local histogram around each feature. The experiments present the improved performance of the overall algorithm over the previous error tolerant matching approaches and Bag-of-visual-words model (BoVW) for object recognition.