Image assets and metadata categorized into single-food crops, landmarks, and street scenes for benchmarking Large Vision-Language Models. The collection focuses on culture mixing scenarios where multiple cultural identifiers co-exist within a single visual frame as defined in the World in a Frame benchmark.
Use Cases
- Benchmark Large Vision-Language Models (LVLMs) using the provided metadata to measure accuracy in multi-cultural cue detection
- Train image classification models to recognize specific cultural artifacts using the SF crop assets
- Analyze model performance on composite scenes containing landmarks and street views to evaluate spatial-cultural reasoning
Strengths
- Includes a dedicated SF directory containing single-food crops rendered on white backgrounds
- Features metadata files that link visual assets to the World in a Frame benchmark categories
- Organizes assets into three primary visual domains: foods, landmarks, and street scenes