We propose a source-free domain adaptation framework that refines pseudo labels using task-domain agreement and adapts CLIP with uncertainty-aware optimization based on proposed Tsallis mutual information.

We proposed novel SKETCH method that extracts high level feature transforming a raw four-dimensional online skeleton point coordinates data to graph images.

We propose an Mixture-of-Experts based continual learning method for Open-Vocabulary Segmentation to handle sequentially collected datasets.