Dewan Tanjil Hossain
M.Sc. Researcher, BUET — Vision-Language Models, Spatio-Temporal & Accessibility-Aware Reasoning
Dhaka, Bangladesh (GMT+6)
M.Sc. researcher at BUET working on vision-language models for spatio-temporal and accessibility-aware reasoning, with a background of 4.5+ years as a software engineer building real-time, high-performance systems. Author of two peer-reviewed papers at ICCIT 2025.
Current Research
Spatio-Temporal Route Reasoning in Vision-Language Models, with an Accessibility-Conditioned Evaluation
Current thesis work building the first benchmark for comparative, multi-frame route reasoning in real outdoor pedestrian environments. Given several candidate routes, each represented as an ordered sequence of street-view frames, a vision-language model must recommend the route best suited to a stated accessibility need (e.g. an elderly person who needs to rest) and justify it with visual evidence that actually exists in the frames.
- RQ1Can current VLMs integrate visual evidence across a multi-frame route sequence and compare alternatives to make a correct, accessibility-appropriate recommendation?
- RQ2When a VLM answers correctly, is it reasoning about the stated need, or answering generically regardless of persona? (tested via a persona-presence ablation)
- RQ3Can the visual justification a VLM gives for its recommendation be trusted? (tested via a structured visual-grounding / hallucination check)
Status: Core dataset of 200 entries across four accessibility personas (aged, general, physical mobility, sensory impairment); evaluation pipeline with dual-judge scoring validated via Cohen’s κ, human-label validation, and Wilson confidence intervals.
Publications
- [1]
REFLECT-R: A Real-World Dataset and Restoration Method for Obstructed Smartphone Lenses
International Conference on Computer and Information Technology (ICCIT), 2025
- [2]
Adoption and Usability of AI Tools in Bangladesh: Opportunities and Challenges
International Conference on Computer and Information Technology (ICCIT), 2025