Meta-Album ACT 40 Mini: Stanford 40 Actions Images for Few-Shot Learning

1,560 images of humans performing 40 actions, preprocessed to 128x128 pixels for the Meta-Album benchmark. The dataset is derived from the Stanford 40 Actions dataset, originally created by researchers including Bangpeng Yao and Li Fei-Fei for action recognition in still images. This mini version was prepared for Meta-Album by Jilin He and released in March 2022.

Use Cases

• Benchmarking few-shot learning algorithms based on the 40 human action classes.
• Training models for human action recognition in still images based on the described image content.
• Developing lightweight computer vision models based on the preprocessed 128x128 pixel image size.

Strengths

• Contains 1,560 images across 40 distinct human action classes, providing a multi-class classification task.
• All images are preprocessed to a consistent 128x128 pixel size, which may simplify model input pipelines.

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.
• The dataset is a mini version; the original Stanford 40 Actions dataset contains 9,532 images, indicating this is a subset.

Provenance

Source

Stanford 40 Actions dataset (http://vision.stanford.edu/Datasets/40actions.html), processed for Meta-Album.

Collection Method

Images were resized to 128x128 pixels using OpenCV as part of the Meta-Album preprocessing pipeline.