Abstract
This paper poses object category detection in images asa type of 2D-to-3D alignment problem, utilizing the largequantities of 3D CAD models that have been made publiclyavailable online. Using the “chair” class as a running ex-ample, we propose an exemplar-based 3D category represen-tation, which can explicitly model chairs of different stylesas well as the large variation in viewpoint. We develop anapproach to establish part-based correspondences between3D CAD models and real photographs. This is achieved by(i) representing each 3D model using a set of view-dependentmid-level visual elements learned from synthesized views ina discriminative fashion, (ii) carefully calibrating the indi-vidual element detectors on a common dataset of negativeimages, and (iii) matching visual elements to the test imageallowing for small mutual deformations but preserving theviewpoint and style constraints. We demonstrate the abilityof our system to align 3D models with 2D objects in the challenging PASCAL VOC images, which depict a wide variety of chairs in complex scenes.