Granular Rigid Body Pose Detection on a Budget

This paper is about structuring an object recognition task for finely controlling computational costs. The motivating problem comes from 3D scans with many identical objects, from which the pose of the objects must be recovered as accurately as possible with a given computational budget. This paper presents a method for detecting such objects which allows control over the amount of computation performed. Using a smaller computational budget results in worse quality results. We also explore different allocations of computation given a particular budget. Given a particular computational budget, different allocations of computation result in different quality results. Devoting computation to the edges of the object results in a tighter fit, at the expense of a narrower basin of attraction. We can achieve the same tightness of fit with roughly half as much computation by focusing on the edges of the object.