Reliable Object Pose Estimation

This thesis addresses estimation of an object’s position and orientation, called the object’s pose, from one or more images. Current methods aim to recover the “best” pose, however, in case of symmetries, occlusions, etc., many poses may explain the observed image, and a single pose estimate cannot explain this ambiguity. Instead of providing a point estimate, we aim to estimate the uncertainty to enable reliability for downstream robot systems.

First, we consider pose estimation for objects on a linear vibratory feeder. Erroneous pose estimates could lead to damaged products or equipment, so reliability is key. We show that the relevant pose uncertainties can be approximated by a distribution over a small discrete set of rotations, and that the uncertainty quantification allows us to reliably avoid failure.

Second, a common approach to pose estimation is establishing correspondences between points in the image and points on the object. Usually, it is assumed that an image point could only correspond to one point on the object, however, that assumption breaks in case of ambiguities such as those imposed by symmetries. We present SurfEmb, modeling continuous, unparameterized correspondence distributions, and we show how to use the distributions to obtain better pose estimates. Our method was on top of the main pose estimation benchmark, BOP, for almost a year.

Third, single-view pose estimation inherently suffers from depth ambiguity and sensitivity to occlusions. We present EpiSurfEmb which optimizes for the pose which maximizes correspondence likelihoods across views. For better pose hypothesis generation, we also combine the image-object correspondence distributions from SurfEmb with epipolar geometry to estimate scene-object correspondence distributions. Our results show, that we can reduce errors by 80-91 %, when multiple images are available.

Fourth, we present Ki-Pode, which flips the correspondence distribution problem to estimating the projection of predefined object keypoints, and we show how the projection-distributions can provide an estimate of the pose distribution. Due to the lack of a way to normalize over the large pose space, we only show distribution estimation on the rotation space, where Ki-Pode provides more reliable estimates across objects than other methods on YCBV.

Fifth, we present SpyroPose which addresses how to scale unparameterized distribution models to pose space. The main idea is to learn distributions at multiple resolutions, allowing more efficient training and many orders of magnitude fewer evaluations at test time. The method can be applied to both rotation and pose space. We present state-of-the-art results on rotation distributions estimation, and on pose distribution estimation, we present the first qualitative results on real images and the first quantitative results at all. We also show that the method extends readily to a multi-view version, presenting a principled way to fuse pose information from multiple images.