On the Use of the Tree Structure of Depth Levels for Comparing 3D Object Views

Abstract

Today the simple availability of 3D sensory data, the evolution of 3D representations, and their application to object recognition and scene analysis tasks promise to improve autonomy and flexibility of robots in several domains. However, there has been little research into what can be gained through the explicit inclusion of the structural relations between parts of objects when quantifying similarity of their shape, and hence for shape-based object category recognition. We propose a Mathematical Morphology inspired hierarchical decomposition of 3D object views into peak components at evenly spaced depth levels, casting the 3D shape similarity problem to a tree of more elementary similarity problems. The matching of these trees of peak components is here compared to matching the individual components through optimal and greedy assignment in a simple feature space, trying to find the maximum-weight-maximal-match assignments. The matching thus achieved provides a metric of total shape similarity between object views. The three matching strategies are evaluated and compared through the category recognition accuracy on objects from a public set of 3D models. It turns out that all three methods yield similar accuracy on the simple features we used, while the greedy method is fastest.

Applicability of Deep Learned vs Traditional Features for Depth Based Classification

“…The subset consists of 10 objects taken from each of 11 object classes and for every object we take four selected views (front, top, top-rear and top-left) for a total of 440 labeled depth images. This setup comes from [4]; some examples of surfaces are shown in figure 1.…”

Component trees for image sequences and streams

“…, T (I n ). This approach provides marginally processing such as in (Dalla Mura et al, 2010) and tree structure comparison (Bracci et al, 2017). The second option is to merge initial trees together and process the final tree.…”

Hypothesis-based Belief Planning for Dexterous Grasping

“…(Kopicki et al 2014). Given an object model, a typical robust global estimator samples subsets of points and computes hypothesised poses based on feature correspondences between the data and the model (Hillenbrand 2008;Tuzel et al 2005;Hillenbrand and Fuchs 2011;Bracci et al 2017). Nonetheless, even for full shape matching, small errors in the pose estimate may lead to critical failures while attempting to grasp, i.e., unexpected contacts may damage the object or the hand itself.…”