Meta AI’s latest innovation, DINOv2, is a self-supervised learning method for training computer vision models that doesn’t require labels or metadata. Unlike traditional image-text pretraining methods, DINOv2 learns to predict relationships between different parts of an image to understand the image’s in-depth information such as spatial relationships and depth estimation. DINOv2 can be used for unsupervised image classification tasks, making it powerful and flexible for businesses’ computer vision applications. DINOv2 could save businesses time and resources, enabling them to develop advanced and sophisticated computer vision applications to improve accuracy, efficiency, and versatility. Real-world applications of DINOv2 include object identification, depth measurement, object classification, object retrieval, and image data curation. In the future, DINOv2 may be integrated with large language models to develop more complex AI systems for businesses.

source update: The Game-Changing Computer Vision AI Model that… – Towards AI