ImageNet-W21-P: A Subset of ImageNet-21K with 10,450 Classes

A subset of the ImageNet-21K (Winter21) dataset, filtered according to the ImageNet-21-P methodology. It contains 10,450 classes, split into training and validation sets. The dataset was processed by re-encoding images into the WEBP format.

Use Cases

• Train image classification models based on the 10,450-class taxonomy.
• Benchmark model performance on a large-scale, multi-class validation split.
• Fine-tune models using the WEBP-encoded image data mentioned in the description.
• Develop techniques for handling large-scale, hierarchical classification tasks.

Strengths

• Contains 10,450 distinct classes, providing a broad classification scope.
• Includes a defined train and validation split for model development and evaluation.
• Images were processed into the WEBP format, which may offer storage or performance benefits.

Limitations

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

Provenance

Source

Filtered from the ImageNet-21K (Winter21) dataset via scripts from Alibaba-MIIL.

Collection Method

Filtering of synsets and re-encoding of images.

Freshness

Last updated 2024-01-07 18:12:01; freshness should be verified.