Project
We explore video language models for understanding human activities, focusing on detailed and fine-grained actions, including the activities of daily living (ADLs) in home environments. We developed InteractADL, a multi-view video dataset with scene graph annotations and atomic action labels to improve computer vision models' ability to understand ADLs. In a series of research projects, we explore novel methods for building and fine-tuning visual language models, which add learnable vectors to dataset class names to better differentiate semantically similar classes.
