Lightweight Activity Understanding

This project develops a lightweight, efficient, and interpretable approach to understanding activities involving multiple people across space and time. Instead of relying on very deep and computationally expensive models, the system focuses on organizing the most important information in a simple and meaningful way. It first looks at each person in the scene and captures information about their appearance, position, and movement over time. It then studies how people interact within each frame and how their actions change across the sequence. These different types of information are combined to understand the overall activity, whether it involves a group of people, two-person interactions, or other multi-entity activity scenarios.