Count Anything

Existing object counting models have been limited in versatility, as they are confined to specific domains such as crowds, vehicles, or cells. To address this, the researchers propose a text-guided object counting approach that takes a natural language query as input, locates the points of targets within an image, and returns their count.

To this end, the researchers built CLOC, a large-scale dataset spanning 6 visual domains (general scenes, remote sensing, histopathology, cell microscopy, agriculture, and microbiology). CLOC includes approximately 220,000 images, 619 categories, and 15 million object instances.

The newly proposed Count Anything model adopts a dual-granularity instance enumeration approach using discrete instance points instead of the conventional density-map method.

• Region-level Sparse Counter: Provides object-level anchors for large, sparsely distributed targets
• Pixel-level Dense Counter: Delivers pixel-wise predictions for small, densely packed, or ambiguously bounded targets
• Complementary Count Fusion: Combines the two counters in a parameter-free manner to maximize accuracy

Experimental results demonstrate that Count Anything surpasses existing open-world counting methods, showing strong accuracy and multi-domain generalization performance.