Viewing Direction Classification: Application to View Planning

An autonomous household robot passively observes the environment while navigating, possibly while performing other tasks, and has spare processing power to identify the different objects it encounters. For example, while setting the table for breakfast, the robot might spot a cup somewhere. Later, when the robot is asked to fetch a cup, it does not need to actively start searching for it, but instead retrieves the cup’s location from memory. We investigated such a passive world-state logging during long-term operation in household environments in the context of object classification [1], [2], [3]. However, manipulating these objects also requires the estimation of object poses. In recent work, we described a method to estimate an object’s orientation using Bingham Mixture Models [5] based on standard local Fast Point Feature Histogram (FPFH) shape features [6] which can deal with high uncertainties and ambiguities. In this work, we focus on analysing which object views and object parts are ambiguous, respective informative, with respect to estimating the object’s orientation in the world. Being able to tell which object views and, more specifically, which object parts are informative is valuable information for active perception (view planning especially in the face of occlusions) as well as passive perception (saving computation by purposively selecting the images to do computation on). Encouraged by the performance improvement and simplicity of the approach to feature selection for categorization by [7], a similar approach, but targeting feature selection for orientation estimation, will be presented and applied to orientation estimation. In our previous work [5], an approach to orientation estimation based on viewing direction classification was presented. This viewing direction classification is motivated by the idea, that the classification allows an insight into which viewing directions and also which object parts are significant for the orientation estimation of the object. This work now investigates this idea for the simulated datasets and shows how a local model of informativeness is obtained using the previously described classification framework.