Despite being a fundamental task in the digital era, Content-Based Image Retrieval (CBIR) faces persistent challenges that limit its widespread adoption. Current methods are often hampered by their reliance on extensive labeled data and meticulous hyper-parameter tuning, rendering them impractical for large-scale applications. We introduce CUSIR (Constraint-based Unsupervised and Scalable Image Retrieval) to overcome these hurdles: a framework engineered to operate without manual supervision or intricate accuracy-tuning. CUSIR constructs an efficient hierarchical index of image features through a top-down, divisive clustering approach. Its core innovation lies in a robust Best-First Search traversal algorithm that ensures retrieval accuracy remains remarkably insensitive to the tree’s structural parameters. This effectively transforms CUSIR into a "set-and-forget" system, eliminating the need for tedious manual configuration to achieve high performance. We benchmarked CUSIR against state-of-the-art methods across five diverse datasets: Caltech-256, Adaptiope, CIFAR-10, Corel-1K, and Corel-10K. Our results reveal that CUSIR, especially when leveraging Vision Transformer (ViT) features, strikes a superior trade-off between mean Average Precision (mAP) and computational speed. With its proven scalability, CUSIR stands as a powerful and practical solution for real-time, large-scale image retrieval tasks.
