Semantic segmentation is a computer vision problem that involves classifying the pixels of an image into meaningful categories. While networks have achieved human-level performance on tasks such as image classification, segmentation performance lags due to the relatively small size of available datasets. In this work, we bypass this problem by training a secondary network to predict when the segmentation network is correct. We hypothesize that the task of recognizing where the segmentation network is wrong is easier than the original task of segmentation. If true, we can use the secondary network to select good predictions from unannotated images. By treating these as ground truth, we can use orders of magnitude more pseudo-training data to train the original network.

Originally, I thought computer science was the most unnecessary field in which to get a PhD. Recently, however, I’ve found that the people doing the most interesting work in machine learning seem to have PhDs. I took the SuperUROP program to get a jumpstart in higher education and research for the fields I’m interested in.