RefCD: Unsupervised Object Detection with Category Awareness

A new approach called Reference-based Category Discovery (RefCD) has been introduced by researchers as an unsupervised object detection technique that enables category-aware detection without the need for manually labeled data. While traditional one-shot detection focuses on closed-set scenarios and relies on labeled datasets, unsupervised methods often produce pseudo boxes that lack category identifiers. RefCD utilizes the similarity of features between predicted objects and unlabeled reference images, incorporating a feature similarity loss to enhance the learning of category-specific characteristics. Additionally, it allows for category-agnostic detection in the absence of reference images, creating a comprehensive framework that addresses the shortcomings of both unsupervised and one-shot methods, facilitating category awareness without incurring annotation costs.