3D Attention-Driven Depth Acquisition for Object Identification

摘要

We address the problem of autonomously exploring unknown objectsrnin a scene by consecutive depth acquisitions. The goal is to reconstructrnthe scene while online identifying the objects from amongrna large collection of 3D shapes. Fine-grained shape identificationrndemands a meticulous series of observations attending to varyingrnviews and parts of the object of interest. Inspired by the recent successrnof attention-based models for 2D recognition, we develop a 3DrnAttention Model that selects the best views to scan from, as well asrnthe most informative regions in each view to focus on, to achieve efficientrnobject recognition. The region-level attention leads to focusdrivenrnfeatures which are quite robust against object occlusion. Thernattention model, trained with the 3D shape collection, encodes therntemporal dependencies among consecutive views with deep recurrentrnnetworks. This facilitates order-aware view planning accountingrnfor robot movement cost. In achieving instance identification,rnthe shape collection is organized into a hierarchy, associated withrnpre-trained hierarchical classifiers. The effectiveness of our methodrnis demonstrated on an autonomous robot (PR) that explores a scenernand identifies the objects to construct a 3D scene model.