Time: 10:00 am
Place: CSB 632
Speaker: Andy Learn
Title: A Perceptual Architecture for Recognizing Spatial Relations
Conducting useful human-computer dialogue about a physical environment (rather than a virtual world) will require the ability to recognize and talk about spatial relationships between objects. In English, we express spatial relations mainly through prepositions (e.g. in, on, above, behind, beside, across, through, etc.) This work seeks to provide a framework for perceptually grounding spatial prepositions. To date, attempts to recognize prepositions using computer vision have succeeded only in highly constrained situations.
I will discuss my proposal for recognizing relations in a real world 3-D environment using an interaction-based approach. I believe that humans' concepts of relations exist to address the needs of motor programs, such as grasping, navigation, etc. By seeking critical visual parameters needed to identify interactive situations and plan actions, I believe we can identify a small, high-leverage feature space that does not require 3-D reconstruction. I envision an architecture for recognizing relations that combines these features in a probabilistic framework that is consistent with the development of spatial relations in children. The architecture is intended to be modular, efficient, and flexible enough to account for differences between languages.