User-Centered Machine Learning for Visual Search

Introduces Visual Search as an augmentation to UCML, enabling users to define custom object detection capabilities in real-time without pre-trained models or data labeling. The system leverages foundational vision-language models to extract features from user-indicated pixels and text prompts, then uses similarity metrics to generate heatmaps and bounding boxes for identified objects across images. This allows for on-demand, ad-hoc object detection for novel or emergent object types.