Arnaud Poletto

This research introduces STAMP-GNN, a novel Graph Neural Network framework that revolutionizes how we understand human visual engagement with dynamic window views. By combining cutting-edge eye-tracking data with computational analysis, the study provides unprecedented insights into occupant behavior and visual attention patterns over time.

• • Graph Neural Network architecture for temporal modeling
• • Integration of eye-tracking data with computational saliency prediction
• • Analysis of visual attention patterns in dynamic window environments

• • Semantic objects account for 57.3% of visual attention, with people attracting 6x more fixations than random patterns
• • Achieved 0.869 AUC on SALICON and +18.6% improvement on video saliency prediction
• • Provided quantitative tools for evidence-based architectural design decisions

This research introduces a novel comparative methodology that systematically evaluates Low Dynamic Range (LDR) and High Dynamic Range (HDR) imaging capabilities for urban daylight analysis. By leveraging deep learning-based segmentation and global video metrics, the study bridges the gap between highly detailed HDR techniques and accessible LDR methodologies for precise lighting analysis.

• • Comparative analysis framework for LDR and HDR imaging techniques
• • Deep learning-based segmentation for automated sky characterization
• • Global video metrics for temporal daylight dynamics analysis
• • Systematic evaluation across diverse urban sky conditions

• • Novel comparative framework for LDR vs HDR imaging evaluation across diverse urban conditions
• • Identified consistent brightness trends under clear skies and discrepancies in complex lighting
• • Demonstrated LDR's critical role in bridging accessible and precise lighting analysis methodologies