Image Caption Pairs for People Holding Objects

2023 images were pulled from Pexels, primarily depicting people holding objects. The dataset includes full images paired with captions generated by the CogVLM model. It was created by author 'lodestones' and last updated on the platform in June 2024.

Use Cases

• Train or evaluate vision-language models using the image files and their corresponding CogVLM-generated captions.
• Fine-tune models for specific scene understanding tasks, such as recognizing human-object interactions like 'person holding a cup'.
• Use image filenames as alternative text labels for captioning or retrieval experiments.
• Analyze the performance of automated captioning systems by comparing model outputs against the provided CogVLM captions.

Strengths

• Captions were programmatically generated using the CogVLM model, ensuring consistency.
• Data source is Pexels, a platform providing professionally shot stock photography.

Limitations

• Specific counts for images, rows, or columns are unknown, making scale assessment difficult.
• Dataset content is biased towards a single theme ('people holding things'), limiting generalizability.
• Image source and caption model are from 2023, which may not represent the latest visual trends or SOTA captioning.

Provenance

Source

Pexels

Collection Method

Images pulled from Pexels; captions generated with CogVLM.

Time Range

2023

Freshness

null

Geography

null