Title: CountGD++: Generalized Prompting for Open-World Counting

Abstract: Current approaches for the automated counting of objects in images and videos are constrained by the rigid methods used to define those objects. Although present techniques permit users to identify targets via text descriptions and visual samples, these visual examples require manual annotation within the image, and the systems lack the ability to exclude specific items from the count. To overcome these limitations, we propose new functionalities that broaden the scope of object specification. Our approach enhances the prompting mechanism to allow users to describe what should be excluded using text and/or visual references. We also introduce "pseudo-exemplars," a technique that automates the generation of visual examples during the inference stage. Furthermore, our counting models are adapted to process visual examples sourced from both real-world and synthetic external images. Additionally, we integrate our novel CountGD++ model as a vision expert agent within Large Language Models (LLMs). Collectively, these advancements significantly boost the flexibility of prompts for multi-modal open-world counting, resulting in marked gains in accuracy, efficiency, and generalization performance across various datasets. The source code can be accessed at https://github.com/niki-amini-naieni/CountGDPlusPlus.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC