RefCOCO-M: Refined Referring Expression Segmentation Benchmark

A cleaned version of the RefCOCO (UNC) validation split for referring expression segmentation. The dataset replaces original instance masks to address poor mask quality and harmful referring expressions, which made the original benchmark an imprecise measure for modern models. It was created by author 'moondream' and last updated on 2025-11-17.

Use Cases

• Benchmarking segmentation model accuracy based on refined ground-truth masks.
• Training vision-language models for referring expression comprehension based on image-text pairs.
• Evaluating the impact of annotation quality on model performance for segmentation tasks.

Strengths

• Addresses known issues of poor mask quality and harmful expressions in the original RefCOCO benchmark.
• Last updated on 2025-11-17, indicating recent maintenance.

Limitations

• Dataset size, row count, and specific file formats are unknown, limiting suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The description is a summary; full details require visiting the external dataset page.

Provenance

Source

huggingface

Collection Method

Derived from the RefCOCO (UNC) validation split with refined instance masks.

Freshness

Last updated 2025-11-17 22:43:51.