Research
My research interests combine artificial intelligence, machine learning, and computer vision. I am particularly interested in understanding and modeling the subtleties of human perception in order to improve accessibility for people with visual impairments.
Projects
Facial Attributes
I have been working to build robust facial attribute models as part of my Ph.D. dissertation. With only one large-scale attribute-labeled dataset to work with, my focus has been on learning models on ideal data that can actually perform well on real-world data.
I have labeled a new dataset: University of Maryland Attribute Evaluation Dataset (UMD-AED). This dataset is sparsely labeled with binary facial attributes and is balanced, with roughly 50 positive and 50 negative labels for each attribute. I collected this data to allow for better evaluation of facial attribute models.
Using Weakly Labeled Data in Video
As part of my work in buildling robust facial attribute models, I started to focus on video and how I could take advantage of weakly labeled data.
No labeled video data was available for attribute prediction, and so I labeled frames from a standard video dataset for face verification: YouTube Faces (Wolf et al. 2011). I labeled four frames from each video in the dataset (first, two middle, last) to evaluate the stability of attribute prediction methods over time. The labels are available here.