Mariya I. Vasileva

My name is Mariya, and I'm interested in all things computer vision, generative models, video analysis and understanding, and ethics of AI systems at scale. I spent a number of years in industry research and worked on a range of problems, from image and video generative models, to visual recommender engines, to 2D-to-3D human body shape and pose modelling, to synthetic data generation for efficient training of foundation models at scale, to fairness and explainability of AI systems. I obtained my PhD in Computer Science from the University of Illinois at Urbana-Champaign under the advisorship of professor David A. Forsyth, where I researched problems in vision and language, visual search and retrieval, and applications of computer vision in the fashion domain.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

Research

My experience lies at the crossover between generative models, multimodal learning, vision and language, foundation models, safety and alignment.

HandsOff is a framework capable of producing synthetic images with corresponding pixel-level labels without requiring additional human annotations. HandsOff utilizes a small number of real labeled images and exploits the rich latent representations of GANs to train an effectively infinite label generator.

This approach learns an image embedding that captures both item similarity and outfit compatibility, respecting item type across a high-level taxonomy. It leverages an end-to-end model to jointly reason about interchangeable items and stylistically compatible combinations. The method is evaluated on 68,306 user-created Polyvore outfits, achieving 3–5% improvement over previous state-of-the-art on outfit compatibility prediction and fill-in-the-blank tasks, while enabling a variety of practical fashion queries.