THIS PAGE INCORPORATES CASCADING STYLE SHEET! TO VIEW PAGE CORRECTLY PLEASE UPDATE TO A CSS COMPLIANT BROWSER
Perceptual Basis of Spatial Relations
UR > CS > Research > Vision > Projects > Perceptual Basis of Spatial Relations
Perceptual Basis of Spatial Relations icon
We seek to visually recognize the spatial relations between objects in a 3-D realistic environment, using an interaction-based approach that identifies a small high-leverage set of visual features.

Andrew Learn (Ph.D. Thesis Work) Randal Nelson (Thesis Advisor) James Allen (Thesis Committee) Chris Brown (Thesis Committee) Brandon Sanders (Research Group) Isaac Green (Research Group)
Perceptual Basis of Spatial Relations figure
Andrew Learn
We want to be able to recognize the spatial relations between objects. e.g. The plane is in the cup, the cup is in front of the monitor, the bear is on top of the can, the car is on the mousepad, the cup/can/mousepad are all on the table, etc. From vision we are extracting key visual features from the objects such as distance, depth, gravity vector, occlusion, surroundedness, etc.

This project seeks to visually recognize the spatial relations between objects in a 3-D, realistic environment. It supports the broader goal of grounding the semantics of natural language in vision. We take an embodied, interaction-based approach to this problem. We believe that humans' concepts of relations exist to address the needs of our motor programs, such as grasping, navigation, etc. When people interact with their environment, they learn visual cues that are correlated with recurrent situations requiring particular actions. By carefully considering important interactions, we seek to identify a small set of basic but critical visual features that can be calculated from images of a 3-D real world setting, but that do not require 3-D reconstruction. We are designing a system that combines these critical visual cues in a modular and progressive computational architecture that is consistent with the development of spatial relations in children. The architecture uses a probabilistic framework that allows generation of single or multiple descriptive prepositions. The end goal is a system that will provide flexible, efficient recognition of spatial prepositions in a three-dimensional real world environment.
http://www.cs.rochester.edu/~learn/spatial_relations
spatial intelligence, spatial relations, prepositions, perception, perceptual, grounding, computer vision, Andrew Learn, symbol grounding, language, spatial cognition, embodied, interaction



Active (February 21, 2003)