Skip to content

Dataset · moondream

Cleaned object referral segmentation benchmark

moondream/refcoco-m

An updated RefCOCO validation split in English with corrected masks and cleaned text queries.

What's inside

A reworked version of the popular RefCOCO (UNC) benchmark validation split by the Moondream team. The authors replaced inaccurate initial segmentation masks with higher-quality ones and removed problematic text descriptions. The dataset contains 1,000 to 10,000 image-phrase pairs pointing to specific objects.

What you can do with it

  • Precise evaluation and benchmarking of modern text-queried segmentation models
  • Testing Vision-Language Models (VLM) for spatial reference understanding in images
Languages
English
Size
1K–10K rows
Downloads this month
1.7K
Likes
59

How to load it

from datasets import load_dataset

ds = load_dataset("moondream/refcoco-m")