Data Science | Ph.D. Qualifier Presentation | Xinlu He

Data Science 

Ph.D. Qualifier Presentation 

Xinlu He, Ph.D. Student 

Monday, April 24th, 2023

Time: 10:00 am -11:00am 

Unity Hall 471 (limited seating available)

Zoom Link: https://wpi.zoom.us/j/98098044123

Committee Members: 

Prof. Jacob Whitehill (Advisor), Computer Science & Data Science, WPI  

Prof. Kyumin Lee, Computer Science & Data Science, WPI  

Prof. Reza Zekavat, Computer Science & Data Science, WPI  

Title:

Person Re-identification in School Classrooms

Abstract:

We explore how to train and apply person re-id systems in school classrooms; this can both facilitate educational research (e.g., help to assess which students get more attention from the teacher) and provide real-time to AI-based educational agents with information about who-is-where-when. Contributions & key results: (1) We investigate a Multi Grade Classroom (MGC) video dataset of school classrooms for the person re-id task. (2) We devised a method of choosing references for the gallery based on combining keypoints and detection boxes from Detectron2, which gives a more stable (lower variance) basis to compare different person re-id algorithms. (3) We substantially improved the re-id R1-Accuracy from 76.3% to 82.4% by fine-tuning with MGC classroom data. (4) We characterize the special challenges (e.g., occlusion, variance in aspect ratio, lack of enough dataset) of classrooms compared to more common person re-id scenarios and discuss the impact on accuracy of these factors.