EVALUATION AND IMPROVEMENT OF PRE-PROCESSING TECHNIQUES FOR MULTIPLE-GAZE GEOMETRY RESEARCH

Abstract

Multiple-gaze geometry, a research project from the Interdisciplinary Visual Intelligence Lab (IVI- LAB) at the University of Arizona, aims to gain an understanding of a 3D scene using a view from a single camera. The locations of points of interest in the scene are inferred by utilizing the intersection of the gazes of the people within it. This task requires several smaller pre-processing tasks to be run on the video data; in particular, people must be detected, along with their faces and key facial features. Currently, these pre-processing steps are performed with older code. Since new detection programs have been developed since the start of this project, there is value in examining how these programs can improve tracking results by providing better face and body detections. In this paper, I show the performance of these detectors on datasets that are being used with the multiple-gaze geometry tracker, and then compare the results visually and quantitatively to those from the currently used detection code. Through this research, I found that the program OpenPose can be used to obtain better body detections, but also that none of the software I tested for face detections outperform the one currently in use.