Item

Abstract

Geometric 3D reasoning has received renewed attention recently, in the context of visual scene understanding. The level of geometric detail, however, is typically limited to qualitative or coarse-grained quantitative representations. This is linked to the fact that today's object class detectors are tuned towards robust 2D matching rather than accurate 3D pose estimation, encouraged by 2D bounding box-based benchmarks such as Pascal VOC. In this paper, we therefore revisit ideas from the early days of computer vision, namely, 3D geometric object class representations for recognition. These representations can recover geometrically far more accurate object hypotheses than just 2D bounding boxes, including relative 3D positions of object parts. In combination with recent robust techniques for shape description and inference, our approach outperforms state-of-the-art results in 3D pose estimation, while at the same time improving 2D localization. In a series of experiments, we analyze our approach in detail, and demonstrate novel applications enabled by our geometric object class representation, such as fine-grained categorization of cars according to their 3D geometry and ultra-wide baseline matching.