New Dataset PerSense-D Enables Model-Agnostic Dense Object Segmentation

PerSense-D introduces a specialized dataset for one-shot personalized segmentation in dense and cluttered images—an area underserved by existing datasets like COCO or LVIS. Comprising 717 images across 28 object categories, each with an average of 39 objects per image, PerSense-D supports research in fields from industrial automation to medical imaging. The dataset was curated through keyword-based retrieval, rigorous manual filtering, and a semi-automatic annotation pipeline that refines model-generated masks. It enables fair evaluation of segmentation methods using mIoU metrics without requiring training, paving the way for more accurate, model-agnostic vision-language segmentation in complex environments.