Grefcoco: Generalized Referring Expression Benchmark for CVPR 2023

Grefcoco is a benchmark dataset for Generalized Referring Expression (GREx) tasks, including segmentation, comprehension, and generation, developed by researcher henghuiding and presented at CVPR 2023. It provides a standardized framework for evaluating models on their ability to process natural language queries that may refer to multiple objects or no objects at all within an image.

Use Cases

• Referring Expression Segmentation (GRES) to generate pixel-level masks for multiple targets described in a single prompt
• Referring Expression Comprehension (GREC) to predict bounding boxes for objects identified by natural language
• Referring Expression Generation (GREG) to synthesize descriptive text for specific image regions

Strengths

• Peer-reviewed benchmark published in CVPR 2023 and IJCV 2026
• Supports multi-target and no-target referring expression scenarios
• Integrates three distinct tasks: GRES, GREC, and GREG

Limitations

• Metadata regarding exact record counts and file sizes is missing from the source
• Likely inherits the geographic and category biases of the COCO dataset

Provenance

Source

Hengshuang Huiding (henghuiding), CVPR 2023 / IJCV 2026

Collection Method

Annotated benchmark based on the COCO dataset

Freshness

Last updated November 2025; associated with publications through 2026.