3D Implicit Shape Models Using Ray Based Hough Voting for Furniture Recognition

The recognition of object categories in 3D scenes is still a challenging problem in computer vision. Many state of the art approaches use Implicit Shape Models, as addressed in [8] and [14], to learn shapes of object categories and a probabilistic Hough Space Voting for the detection of instances of the learned category. In this paper we present a novel 3D Hough Space Voting approach for recognizing object categories, learned from artificial 3D models, in 3D scenes. The proposed method uses rays instead of points to vote for object reference points. The usage of ray voting allows a clustering of votes, showing in similar directions, to a single vote with an appropriate vote weight. The main advantage for the Implicit Shape Model is that it can be trained with an unlimited amount of training data, while keeping the upper bound of computation effort constant. In addition, it is also able to abstract from the model sizes which is very helpful when training with artificial models taken from different sources and modelled in different scales. We validate our approach in two tasks: an object categorization is performed on a standard 3D dataset of artificial models and a recognition of furniture categories is evaluated on a dataset of captured indoor room scenes.