count-anything

Why we included this project

Count Anything answers a deceptively simple question: given a picture and a text query like "how many cells are here", it returns a set of points marking each instance, and the number of points is the count. Because the output is grounded in actual locations rather than a bare number, you can see where the model thinks each object is and judge whether to trust it. A single checkpoint covers six visual domains, from general scenes and remote sensing to histopathology, cellular microscopy, agriculture, and microbiology, so one model can replace several domain-specific counters. The repository includes a Hugging Face demo, a downloadable checkpoint, and setup instructions, along with the CLOC benchmark and its v1.1 annotation refresh for reproducing the evaluation. That combination makes it a practical pick for quality inspection, biomedical image analysis, and agricultural monitoring, where counting accuracy and interpretability both matter.