Learning 3-D object orientation from images

Abstract

We propose a learning algorithm for estimating the 3-D orientation of objects. Orientation learning is a difficult problem because the space of orientations is non-Euclidean, and in some cases (such as quaternions) the representation is ambiguous, in that multiple representations exist for the same physical orientation. Learning is further complicated by the fact that most man-made objects exhibit symmetry, so that there are multiple "correct" orientations. In this paper, we propose a new representation for orientations-and a class of learning and inference algorithms using this representation-that allows us to learn orientations for symmetric or asymmetric objects as a function of a single image. We extensively evaluate our algorithm for learning orientations of objects from six categories.

On the Continuity of Rotation Representations in Neural Networks

“…Grassia et al [16] pointed out that Euler angles and quaternions are not suitable for orientation differentiation and integration operations and proposed exponential map as a more robust rotation representation. Saxena et al [28] observed that the Euler angles and quaternions cause learning problems due to discontinuities. However, they did not propose general rotation representations other than direct regression of 3x3 matrices, since they focus on learning representations for objects with specific symmetries.…”

Efficient grasping from RGBD images: Learning using a new rectangle representation

“…Saxena et al [1,2,23] showed that a 'grasping point' could be estimated from the image using supervised learning algorithms, and that this method generalized to a large number of novel objects. However, their grasping point representation only indicated where to grasp, and other parameters such as gripper orientation were left to be estimated by other methods (e.g., [24]). In later works, Saxena et al [25] also used point clouds for grasping.…”

Unstructured human activity detection from RGBD images

“…For 10 joints, we convert each rotation matrix to half-space quaternions in order to more compactly represent the joint's orientation. (A more compact representation would be to use Euler angles, but they suffer from representation problem called gimbal lock [31].) Along with these joint orientations, we would like to know whether person is standing or sitting, and whether or not person is leaning over.…”