In Partial Fulfillment of the Requirements for the Degree of Master of Science

Justin Brown

will defend his thesis

Object Detection using Segmentation

Abstract

Object detectors that are based on bounding box regression are complex and require a lot of refinement to get good results. Is there a simpler way to do object detection? To that end, I present a new method for object detection where you first segment an image, and then cut each object from the segmentation to produce its bounding box. The key to this method is that it uses only ground truth bounding box data to generate the ground truth segmentation mask, rather than using a semantic segmentation mask which are pixel-perfect. Additionally, I present a modified Flood Fill algorithm for the cutting task. Using this method, I eliminate the need for bounding box regression, anchor-boxes, region proposal methods and all of their associated complexities. Experiments show that my method gets competitive performance on the WIDER Face dataset with full size images and runs between 30 and 50 FPS when using 640x480 px images.

Date: Thursday, April 18, 2019
Time: 3:00 PM
Place: PGH 362
Advisors: Dr. Shishir Shah

Faculty, students, and the general public are invited.