Junction assisted 3D pose retrieval of untextured 3D models in monocular images

This article presents a comprehensive framework for the recognition of untextured 3D models in a single image. The method proposed here is capable of recovering a 3D pose in a few hundred of milliseconds, which is a difficult challenge using this type of model.This proposal deals with 3D models that lack texture, so geometry features of the model are used as a basis of the 3D pose retrieval. An automatic process extracts the junctions and contours of the model, replacing the user interaction. Junctions will provide us an efficient mechanism to generate candidate matches, while contours will select the correct match based on a robust shape similarity evaluation. Our method only requires the 3D triangle mesh of the model as input, since the rest of the process is done automatically.We demonstrate the behaviour of our approach against a variety of real scenes and models. Moreover, we explain how to face the first pose problem in a robust way using a history of votes. We also present a study of the method parameterisation, describing the influence of each parameter.

► New 3D recognition method based on geometry.
► All geometric constraints are extracted automatically during an offline process.
► Monocular detection of non-textured 3D models with different environment conditions.
► Wide applicability in industrial environments, where untextured models are quite common.
► It can be used to solve the first pose problem and to recover from a tracking failure.