HalluSegBench is a counterfactual segmentation reasoning dataset for evaluating pixel-grounding hallucination in vision-language segmentation models.
Each example contains a factual image, a factual segmentation mask, a counterfactual image, and a counterfactual segmentation mask. The counterfactual image replaces the original target object or region with a plausible alternative while preserving the surrounding scene context. The dataset is designed to test whether… See the full description on the dataset page:
https://huggingface.co/datasets/PLAN-Lab/HalluSegBench.
