CountingDINO: A Training-free Pipeline for Class-Agnostic Counting using Unsupervised Backbones

CountingDINO represents a breakthrough in class-agnostic counting by eliminating the reliance on labeled data, which has traditionally limited the scalability and generalization of such methods. By employing a fully unsupervised feature extractor, CountingDINO can effectively estimate object counts in images across various categories. Its performance was rigorously evaluated on the FSC-147 benchmark, where it consistently outperformed a baseline based on a state-of-the-art unsupervised object detector. This achievement not only demonstrates the framework's effectiveness but also highlights its potential to transform applications in computer vision, making it a valuable tool for researchers and practitioners alike.