I am a PhD candidate at Department of Electrical Engineering and and Computer Science, York University, working under supervision of James Elder. My PhD research focuses on video understanding for sports. I am a professional software developer with over 15 years of experience. I have received my MSc and BSc in Computer Science from York University in 2003 and 2001.

• Dec 2021 - Completed internship at Vector Institute. Presented my work: MLOps for Healthcare
• Sep 2021 - Joined Vector Institute as Applied AI Intern.
• Jun 2021 - "Contrastive Learning for Sports Video: Unsupervised Player Classification" is a runner-up for Sportlogiq Best Paper Award.
• May 2021 - Presented invited talk "Towards Sports Video Understanding: Player Detection, Classification and Tracking" at Ontario Tech University.
• Apr 2021 - Paper M. Koshkina, H. Pidaparthy, J.H. Elder "Contrastive Learning for Sports Video: Unsupervised Player Classification" accepted to CVPR Vision in Sports Workshop.
• Jan 2021 - Started as Machine Learning Team Lead (part-time) at RPS Canada Inc. (Formerly Geminare).

Detecting and tracking players throughout the game is important for game analysis and automatic stats collection. It supports other sports video understanding tasks such as game event detection. Current state-of-the-art approaches are focused on MOT (Multi-object tracking) of pedestrians and vehicles. Task of tracking players brings its own challenges: players on the same team have very similar appearance, fast motion and frequent occlusions complicate the task. In our research, we are building on top of current MOT state-of-the-art methods while addressing sports-specific challenges.

Jersey number recognition is key task in sports video analysis, due in part to its importance for long-term player tracking. Unfortunately, each player's jersey number is only visible in a fraction of video frames. I am investigating how scene text recognition (STR) methods can be adapted to the problem of jersey number recognition. I also aim to asses the degree to which STR models trained on one sport can generalize to another.

We address the problem of unsupervised classification of players in a team sport according to their team affiliation, when jersey colours and design are not known a priori. We adopt a contrastive learning approach in which an embedding network learns to maximize the distance between representations of players on different teams relative to players on the same team, in a purely unsupervised fashion, without any labelled data. We evaluate the approach using a new hockey dataset and find that it outperforms prior unsupervised approaches by a substantial margin, particularly for real-time application when only a small number of frames are available for unsupervised learning before team assignments must be made. Remarkably, we show that our contrastive method achieves 94% accuracy after unsupervised training on only a single frame, with accuracy rising to 97% within 500 frames (17 seconds of game time). We further demonstrate how accurate team classification allows accurate team-conditional heat maps of player positioning to be computed.

Maria Koshkina, Hemanth Pidaparthy, James H. Elder Contrastive Learning for Sports Video: Unsupervised Player Classification (2021), In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 4528-4536. .
For more info see our project page and code.