LSD-Net: Look, Step and Detect for Joint Navigation and Multi-View Recognition with Deep Reinforcement Learning

N dinesh reddy

Feb 15, 2018 (modified: Feb 15, 2018) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Multi-view recognition is the task of classifying an object from multi-view image sequences. Instead of using a single-view for classification, humans generally navigate around a target object to learn its multi-view representation. Motivated by this human behavior, the next best view can be learned by combining object recognition with navigation in complex environments. Since deep reinforcement learning has proven successful in navigation tasks, we propose a novel multi-task reinforcement learning framework for joint multi-view recognition and navigation. Our method uses a hierarchical action space for multi-task reinforcement learning. The framework was evaluated with an environment created from the ModelNet40 dataset. Our results show improvements on object recognition and demonstrate human-like behavior on navigation.