This repository presents a zero-shot text-guided object counting framework enhanced with exemplar-based learning. The system is designed to improve counting accuracy by explicitly modeling count-relevant visual features and suppressing background interference.
- Integrated a loss function that explicitly differentiates foreground objects from background regions, enabling the model to focus on count-relevant features.
- Developed an exemplar extraction module to identify representative object instances from images and leverage these features during both training and evaluation.
Demo videos illustrating the counting performance and exemplar-guided inference are available at here