Ma, YezhouWang, Kuan-Wen2020-08-102020-08-102020-05Ma_cornell_0058O_10832http://dissertations.umi.com/cornell:10832Wang_cornell_0058O_10834http://dissertations.umi.com/cornell:10834https://hdl.handle.net/1813/7021831 pagesIn this study, we explore analyze the VizWiz dataset in a computer vision perspective, and identify challenges and potential solutions toward building computer vision model for visually impaired people. We conducted analysis on the images and question-answer(QA) pairs on the VizWiz dataset to identify common domains of problems. We also targeted object detection as a first step. By analyzing and mapping the QA pairs to ImageNet labels, we found that building a new set of labels specifically designed for this domain would be crucial. We then inspect and build a vocabulary set for the object detection task.endissertation or thesishttps://doi.org/10.7298/5mg1-s592