2D object detection is the task of finding (i) what objects are present in an image and (ii) where they are located, while 3D pose estimation is the task of finding the pose of these objects in 3D space. State-of-the-art methods for solving these tasks follow a two-stage approach where a 3D pose estimation system is applied to bounding boxes (with associated category labels) returned by a 2D detection method. This paper addresses the task of joint object category and 3D pose estimation given a 2D bounding box. We design a residual network based architecture to solve these two seemingly orthogonal tasks with new category-dependent pose outputs and loss functions, and show state-of-the-art performance on the challenging Pascal3D+ dataset.
View on arXiv