Label Mappings for Popular Computer Vision and NLP Datasets

This repository provides label mapping files that convert integer IDs to human-readable class names for multiple datasets. It includes mappings for ImageNet-1k, ImageNet-22k (with 21,843 classes), COCO detection 2017, and others like VQAv2 and Kinetics-700. The files are maintained by HuggingFace and were last updated in June 2025.

Use Cases

• Decode integer predictions from models trained on ImageNet-22k using the provided id2label mapping for its 21,843 classes.
• Map COCO detection 2017 class IDs to object names for interpreting bounding box detection outputs.
• Translate VQAv2 answer IDs to textual answers for evaluating visual question answering model performance.
• Convert Kinetics-700 action IDs to human-readable action names for video classification tasks.

Strengths

• Covers multiple major datasets including ImageNet-22k, which has 21,843 distinct classes.
• Includes label files for both computer vision (e.g., COCO, Cityscapes) and NLP/VQA (e.g., VQAv2) tasks.
• Maintained by HuggingFace with a recent update in June 2025.

Limitations

• The specific file formats, sizes, and exact structure of the mapping files are not detailed in the provided input.
• The repository's scope is limited to providing label mappings and does not contain the actual dataset images, videos, or annotations.

Provenance

Source

HuggingFace

Collection Method

Repository of label mapping files (id2label) compiled for use with HuggingFace Transformers.

Freshness

Last updated on 2025-06-12.