Viewpoint detection models for sequential embodied object category recognition

Abstract

Abstract-This paper proposes a method for learning viewpoint detection models for object categories that facilitate sequential object category recognition and viewpoint planning. We have examined such models for several state-of-the-art object detection methods. Our learning procedure has been evaluated using an exhaustive multiview category database recently collected for multiview category recognition research. Our approach has been evaluated on a simulator that is based on real images that have previously been collected. Simulation results verify that our viewpoint planning approach requires fewer viewpoints for confident recognition. Finally, we illustrate the applicability of our method as a component of a completely autonomous visual recognition platform that has previously been demonstrated in an object category recognition competition.

A Portable Active Binocular Robot Vision Architecture for Scene Exploration

“…By carrying out a qualitative comparison with current robot vision systems whose performance has been reported in the literature, we argue that our architecture clearly advances the reported state-of-the-art [5,18,15,3,13] in terms of our system's innate visual capabilities and portability to different environment settings, e.g. multiple same-class object identification and tolerated degree of visual scene complexity.…”

“…Should a visual task becomes ill-posed, the gaze of a robot can be shifted to perceive the scene from a different viewpoint [7]; and therefore a better understanding of the task. Current research in active robot heads has focused on the "lost and found" problem [15]. That is, a robot is commanded to search and locate an object in its working environment for exploration tasks [10,6], manipulation tasks [18,20] and/or navigation [15].…”

“…[18,15,13]). These heuristic approaches are mainly driven by the outputs of available feature extraction techniques.…”

“…For example, Rasolzadeh et al [18] used depth to segment the scene according to the distance between a targeted object and the robot as part of a visual object search heuristic. Likewise, Merger et al [15] implemented a saliency map that combines intensity, colour and depth features to drive attention, biased by a top-down feature detection based on the MSER feature extractor [14] for object recognition and navigation. Aydemir et al [6] have recently presented a strong correlation between local 3D structure and object placement in everyday scenes.…”

Minimum-time trajectories for kinematic mobile robots and other planar rigid bodies with finite control sets

“…Laporte et al [10] is one particularly similar example where a Bayesian formulation is used. Even more similar are approaches targeted to recognizing objects in indoor scenes, such as [11] and our previous work [12], [13]. We have previously proposed the use of a generative model to combine detections, but this is our first attempt to include occlusion reasoning as a variable within the model, which increases the performance in high clutter.…”