April 14, 2008
ECE graduate student Devi Parikh and Professor Tsuhan Chen won a Best Paper Award at the 2007 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Beyond Patches Workshop. The paper titled "Unsupervised Learning of Hierarchical Semantics of Objects (hSOs)" proposes an algorithm to determine higher-level semantic relationships among objects in a scene such as an office, which is represented via a hierarchical structure.
"Consider a robot looking for a keyboard in an office to perform some automated task," says Parikh. "However, say the keyboard is not clearly visible due to occlusions in the scene such as the presence of a human at the desk blocking most of the view of the keyboard, but the robot can clearly spot the computer monitor. One can imagine that having detected the monitor, the task of finding the keyboard can be made much more accurate and efficient, if only the robot could understand the semantic relationship between the monitor and keyboard as we do."
In their paper, Parikh and Chen present an algorithm that enables a machine/robot to learn such semantic relationships automatically. The algorithm infers these semantics by analyzing a collection of images of a scene taken over a period of time and building a statistical model of the locations of the objects with respect to each other. A key characteristic of the algorithm is that it doesn't require any annotations of the images by humans and is thus fully automatic. Apart from providing contextual information for accurate localization and recognition of objects in a scene, other potential applications of the proposed work are anomaly detection and scene recognition.