Modeling pose/appearance relations for improved object localization and pose estimation in 2D images

Abstract

We propose a multiview model of appearance of objects that explicitly represents their variations of appearance with respect to their 3D pose. This results in a probabilistic, generative model capable of precisely synthesizing novel views of the learned object in arbitrary poses, not limited to the discrete set of trained viewpoints. We show how to use this model on the task of localization and full pose estimation in 2D images, which benefits from its particular capabilities in two ways. First, the generative model is used to improve the precision of the pose estimate much beyond nearest-neighbour matching with training views. Second, the pose/appearance relations stored within the model are used to resolve ambiguous test cases (e.g. an object facing towards/away from the camera). Here, changes of appearance as a function of incremental pose changes are detected in the test scene, using a pair or triple of views, and are then matched with those stored in the model. We demonstrate the effectiveness of this method on several datasets of very different nature, and show results superior to state-of-the-art methods in terms of accuracy. The pose estimation of textureless objects in cluttered scenes also benefits from the proposed contributions.