MG-Data: 8M Images with Multi-Granularity Vision-Language Annotations

8M images sourced from CC3M and CC12M, each annotated with a short caption, a long caption averaging over 100 words, and pixel-level masks. SEU-VIPGroup created this dataset for fine-grained vision-language alignment, containing over 40 million object masks and 25 million region masks.

Use Cases

• Training models for detailed image captioning based on the long textual descriptions averaging over 100 words.
• Developing fine-grained object detection systems based on the 40 million+ pixel-level object masks.
• Building models for region-level visual understanding based on the 25 million+ region masks.
• Benchmarking vision-language alignment models on multi-granularity annotations spanning image, object, and region levels.

Strengths

• 8M images provide a large-scale foundation for training.
• Multi-granularity annotations include image-level captions, object-level masks (40M+), and region-level masks (25M+).
• Long captions averaging over 100 words offer detailed textual descriptions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the CC3M and CC12M collections.

Provenance

Source

Images sourced from CC3M and CC12M collections.

Collection Method

Annotated with short captions, long captions, and pixel-level masks.

Freshness

Last updated 2026-06-06 09:47:15; freshness should be verified.