COCO-AB: Extended Annotations for 82,765 COCO 2014 Images

COCO-AB is an extension of the COCO 2014 training set, enriched with additional annotation byproducts. The dataset contains 82,765 reannotated images from the original set. It was authored by coallaoh and last updated on Hugging Face in July 2023.

Use Cases

• Training object detection models based on the reannotated bounding boxes.
• Analyzing scene interactions based on the recorded additional actions.
• Benchmarking model performance on an enriched version of the COCO 2014 training set.
• Studying annotation consistency and quality across different labeling efforts.

Strengths

• Contains 82,765 images, providing a substantial scale for model training.
• Builds upon the widely-used COCO 2014 benchmark, ensuring relevance.
• Includes additional annotations aimed at providing a richer understanding of images.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Extension of the COCO 2014 training set.

Collection Method

Reannotation of the original images to record additional actions and interactions.

Time Range

Based on COCO 2014 data.

Freshness

Last updated 2023-07 23 18:22:22; freshness should be verified.

Geography

Likely follows the geographic distribution of the original COCO dataset.