Exploratory Research in Scene Analysis and Object Recognition

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Abstract In this SGER proposal, the PIs propose to develop novel tools that allow qualitative 3D vision from 2D images and video sequences. Most of today's approaches to visual object recognition essentially reduce this problem to one of pattern classification, where rectangular image patches are independently compared to stored templates to produce isolated object labels. The proposed research explores new research directions for the task of recovering the 3D layout of a scene from a single image and for using the 3D layout to help in recognizing object categories in a scene. Each of the research directions proposed for exploration has the potential of opening up an entire new set of approaches and algorithms and has the potential of defining an entire new field of Computer Vision, which the PIs call "qualitative geometric reasoning", as opposed to the traditional quantitative approaches which assume precise depth and dense measurements from stereo or SFM. By advocating the use of qualitative geometric reasoning, this body of work is expected to contribute to a radical change in the way the image interpretation and scene analysis problems are tackled in the computer vision community. The proposed research is anticipated to result in new directions in the general area of geometric reasoning for scene analysis, which is a critical enabling technology for a wide range of applications including defense, health care, human-computer interaction, image retrieval and data mining, industrial and personal robotics, manufacturing, scientific image analysis, space exploration, surveillance and security, and transportation.