FURI | Fall 2025

Advancing Video Question Answering (VideoQA) Through Attribution, Reasoning, and Counterfactual Inference Tasks

This research investigates how to equip video question answering (VideoQA) systems, AI models that answer questions about videos, with deeper reasoning for real-world scenarios. Current systems handle surface-level prompts but struggle to attribute causes, count objects, reason over events, infer prior states, or explore counterfactuals. The project will curate and annotate urban video data, fine-tune transformer-based models on these five tasks, and design evaluation and explainability methods. The outcome aims at safer, more interpretable systems for traffic monitoring and related domains.