An indexing scheme for efficient data-driven verification of 3D pose hypotheses

摘要

We present a novel approach for efficient verification of 3D pose hypotheses using 3D sensory data. This task is generally performed by exhaustively comparing scene and model features at a high computational cost. We avoid exhaustive feature comparison by using a highly discriminative indexing scheme, which can significantly improve the efficiency of the verification task. The proposed scheme is data-driven, i.e. it is used to determine model features consistent with a given scene feature. For a specific verification task, a number of generic scene features are considered (e.g. surface patch, edge segment). For each generic feature, we compute a set of subspaces in the 6D pose space; each one of them bounds object poses that lead to a consistent match between the scene feature and a model feature. These subspaces are made invariant to the location and orientation of the scene feature in the world coordinate frame by expressing them with respect to a data-centered coordinate frame. This enables us to index them off-line, which is done using a multi-dimensional data structure. Given a pose hypothesis to be verified at run-time, model features consistent with a scene feature are directly accessed through navigating the appropriate index using parameters of the hypothesis as keys. Time complexity of the proposed scheme is analyzed. Its performance is demonstrated using a system for localizing a polyhedral object in a robot hand using vision and touch.