The Breakfast actions dataset Official

A dataset of breakfast preparation activities performed by 52 different individuals in 18 different kitchens, designed to reflect real-world recognition scenarios.

Introduced at CVPR 2014, this dataset comprises 10 breakfast preparation activities and supports activity recognition and temporal segmentation in everyday settings.

Dataset specifications: ~77 hours of video (>4 million frames), 320×240 pixels resolution (down-sampled), 15 fps frame rate, 3-5 uncalibrated cameras per location.

Downloads

Download Breakfast data and annotations on Hugging Face. The dataset page lists the available archives, their contents, checksums, and links to additional features.

The reduced Fisher-vector features and coarse and fine segmentation annotations are available there. The raw video mirror is incomplete: the original Google Drive download exceeded its quota when checked on October 11, 2026. See the dataset page for video availability.

HTK action-recognition code and evaluation instructions describe the feature format and training protocol.

Action classes (10)

Current benchmarks

Fully supervised learning

  • Activity Classification: Classify full videos according to the 10 activity classes
  • Temporal Segmentation: Detect and classify the 48 action units within the videos

Weakly supervised learning

  • Temporal Alignment: Detect actions given transcripts at test time
  • Temporal Segmentation: Detect and classify 48 action units without transcripts

Train/test splits

The dataset includes predefined splits for evaluation:

Citation

Please cite the following papers when using this dataset:

Primary paper:
H. Kuehne, A. B. Arslan and T. Serre. The Language of Actions: Recovering the Syntax and Semantics of Goal-Directed Human Activities. CVPR, 2014.

Follow-up work:
H. Kuehne, J. Gall and T. Serre. An end-to-end generative framework for video segmentation and recognition. WACV, 2016.