Speech theme:

Embodied Vision and SLAM: An Online Learning Approach

Speech summary:

The ultimate purpose of embodied intelligence is to realize multi-level closed-loop processing among agents and their bodies, environments, and actions, and thus makes the intelligent machines being tightly embedded in environments. The research of SLAM aims to develop machine perception systems that can accomplish simultaneous localization and mapping of mobile robots, and it is a core and indispensable technique for the embodied intelligence. In order to enhance generalization ability of an embodied vision system in real situations, we have to maintain a high adaptability of vision systems to variations of environments and tasks, and online learning method provide an effective approach to the end. To demonstrate current trends in the online learning based SLAM research, the talk will report some new progresses in the field, including: a dynamic SLAM method taking advantage of forgetting mechanism of online learning methods; geometric and instance reconstruction of 3D environments by using monocular videos.