Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 01, 2022 - June 04, 2022
Training a domain-invariant viewpoint planner (VP) is an important task in visual place recognition tasks for autonomous mobile robots. robots. The objective of this problem is to acquire effective viewpoint shifts (e.g., landmark observation behavior) for place recognition based on visual experience under past domains (e.g., season, weather, illumination, etc.). Existing VP techniques assume that the domain is almost invariant, and none of them has ever dealt with essential cross-domains. In order to train domain-invariant VPs, we need domain-invariant scene representations. Inspired by the recent emergence of techniques for training graph classifiers (e.g., graph convolutional networks), we focus on research and development of new domain-invariant scene representations based on semantic scene graphs.
