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 dataset was created by PLAN-Lab and was last updated on 2026-06-21.
Use Cases
- Benchmarking model hallucination based on the counterfactual image-mask pairs
- Evaluating pixel-grounding accuracy based on the comparison between factual and counterfactual segmentation masks
- Training models for robust scene understanding based on the plausible alternative object replacements
Strengths
- Designed specifically for evaluating pixel-grounding hallucination, a targeted use case.
- Contains paired factual and counterfactual images with corresponding segmentation masks.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- PLAN-Lab
- Freshness
- Last updated 2026-06-21 04:26:55; freshness should be verified.