Philippos Mordohai, George Kamberov and Gerda Kamberova (Hofstra University) received a Google Research Award for the amount of $50k to develop novel computer vision techniques for scene understanding in large-scale urban environments. Data will be collected by a sensor suite acquired with funding provided by the National Science Foundation (NSF) under a Computing Research Infrastructure award, which includes state of the art navigation equipment, as well as an omnidirectional LIDAR sensor, an omnidirectional video camera and other narrow field of view cameras.
The proposed approach takes advantage of the multiple observations captured from each object as a vehicle carrying the sensors traverses the scene. The factors that distinguish this effort from conventional approaches is that processing is driven by 3D information, since segmentation is usually easier in 3D, aided by image-based properties. Novel invariant descriptors that capture both geometry and appearance will be used for object recognition within a scene understanding framework that exploits contextual cues, such as roads, buildings and object co-occurrence probabilities. A strong emphasis will be placed on recognizing significantly more object and surface categories than current approaches, emphasizing functionally significant categories, such as entrances, and dynamic objects.
Copyright © 2017 Stevens Institute of Technology. All Rights Reserved.