CSK-Detector: Commonsense in object detection

摘要

We propose an approach CSK-Detector for object detection and image categorization, well-suited for big data, by transferring commonsense knowledge from a knowledge base, augmented with premises and quantifiers. It is implemented for domestic robotics, especially with the motivation that next-generation and multipurpose domestic robots should be able to seamlessly discern environments for specific tasks without prior annotation of excessive images. CSK-Detector is evaluated on real data, yielding better results than deep learning without commonsense, while also providing an explainable approach. It broadly impacts human-robot collaboration and smart living.