Visual Diversity Budget: Geo-Data Similarity for Optimized Annotation

MichiganNLP released the Visual Diversity Budget dataset in May 2024 to optimize the balance between model performance and annotation costs using geo-data similarity. The dataset provides resources for multimodal deep learning research, specifically targeting geo-diverse visual data for models like BLIP-2 and CLIP.

Use Cases

• Reducing annotation costs by filtering data based on geo-data similarity metrics
• Benchmarking BLIP-2 model performance across geo-diverse visual samples
• Conducting diversity analysis on multimodal datasets to identify geographic representation gaps

Strengths

• Provides geo-diverse metadata for multimodal model alignment
• Includes specific similarity metrics for CLIP and BLIP-2 architectures
• Maintained by the MichiganNLP research organization

Limitations

• Lack of documented column headers or schema in the primary metadata
• Total record count and file size are not specified

Provenance

Source

MichiganNLP GitHub repository

Collection Method

Annotated geo-data similarity analysis

Freshness

Last updated May 20, 2024.

Geography

Global