Embedding geometry in generative models for pose estimation of object categories

Abstract

Regression-based models built on local gradient-based feature descriptors have showed to be successful for continuous pose estimation of object categories. Nonetheless, a crucial weakness of these methods is that no geometric information is taken into account. Therefore, geometrically inconsistent poses may be preferred, and this forces to employ a coarse-grained pose estimator as a pre-processing step to avoid potentially large estimation errors. In this paper, we propose a method that combines generative feature models and graph matching techniques in a unified probabilistic formulation of the continuous pose estimation problem. Our approach retains the lightness and generality of generative feature modeling, while favoring geometrically consistent results. Experiments show that pose pre-processing steps are not needed if geometry is embedded in the matching stage. We evaluated our approach on two different car datasets and we experimentally show that our algorithm outperforms state-of-the-art methods by 25%.