Candidate: Danial Sadrian Zadeh
Date: July 31, 2026
Time: 2:00 PM
Location: Online
Supervisor: Otman A. Basir
Co-Supervisor: Behzad Moshiri
All are welcome!
Abstract:
Contextual traffic scene understanding encompasses key aspects of safe navigation and decision-making in autonomous vehicles, which mainly involve perceiving and understanding the traffic environment to achieve situational awareness by capturing dynamic traffic patterns. In this context, this seminar focuses on three proposed strategies that leverage the advantage signal in Self-Critical Sequence Training (SCST) to improve the quality of generated captions within the driving scene paradigm. These methods are five advantage-scaling methods, one advantage-fusion method based on a novel fusion rule, and three sampling-temperature adaptation methods. The effectiveness of a Large Language Model (LLM) as an evaluation metric is also evaluated for fine-tuning and validation. Initial experiments and results show that most of the proposed hypotheses increase metric scores, such as CIDEr-D and SPICE.