GSEval is a benchmark dataset containing 3,800 images, curated by hustvl and last updated on 2025-05-20. It is designed to evaluate the performance of AI systems in pixel-level and bounding box-level grounding based on natural language descriptions.
Use Cases
- Benchmarking pixel-level grounding models based on the described image collection
- Evaluating bounding box-level grounding performance based on natural language prompts
- Assessing AI system capabilities in understanding and localizing objects in images
Strengths
- Contains 3,800 images specifically curated for evaluation
- Designed for assessing both pixel-level and bounding box-level grounding tasks
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
Provenance
- Source
- hustvl
- Freshness
- Last updated 2025-05-20 03:57:53; freshness should be verified