Datasets:
The Breakfast Actions Dataset
Breakfast contains recordings of 52 people preparing breakfast in 18 kitchens: approximately 77 hours of video, with 10 activities and 48 action units. It was introduced at CVPR 2014 for studying activity recognition and temporal segmentation in everyday settings.
Downloads
The archives below retain the original file formats and directory structures. Download and unpack them directly; this repository is an archive distribution, not a table for the Hugging Face Dataset Viewer.
| File | Contents | Size |
|---|---|---|
| breakfast_data.tar.gz | Reduced Fisher-vector features for the HTK training and test demo; these are not raw videos or segmentation annotations | 998,401,265 bytes |
| segmentation_coarse.tar.gz | Coarse temporal annotations in XML and text formats | 291,172 bytes |
| segmentation_fine.tar.gz | Fine temporal annotations in XML and text formats | 931,312 bytes |
| demo_bundle.rar | Legacy MATLAB demo | 71,527 bytes |
Raw videos: the original Google Drive archive contains the 320×240, 15 fps release (3,930,212,562 bytes). On October 11, 2026, Google Drive blocked the full download because its download quota was exceeded. The video archive is therefore not mirrored here yet. Google advises trying again later; a valid video copy is needed before this mirror can be completed.
Additional features: I3D RGB and flow features remain on Google Drive. The four dense-trajectory archives remain on Dropbox: split 1, split 2, split 3, and split 4. These external archives have not been mirrored or fully validated as part of this repair.
Archive repair and checksums
On October 11, 2026, three HTML error pages incorrectly stored as .tar.gz files were replaced with the original feature and segmentation archives. Each replacement passed a complete gzip integrity check and tar-directory inspection. The invalid BreakfastII_15fps_qvga_sync.tar.gz placeholder was removed; it was not a usable video archive.
SHA256SUMS provides checksums for the three restored archives. Archive provenance records their source URLs and validation results.
The feature source is linked from Hilde Kuehne's project page. The annotation sources are the original downloads listed on the lab dataset page.
Code and evaluation
The HTK action-recognition repository documents the reduced feature representation, training and evaluation. Its README explains how to use breakfast_data.tar.gz. Follow the published evaluation protocol when comparing results; feature files and video clips are different representations of the recordings.
Citation
@inproceedings{kuehne2014language,
title={The language of actions: Recovering the syntax and semantics of goal-directed human activities},
author={Kuehne, Hilde and Arslan, Ali and Serre, Thomas},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
year={2014}
}
@inproceedings{kuehne2016end,
title={An end-to-end generative framework for video segmentation and recognition},
author={Kuehne, Hilde and Gall, Juergen and Serre, Thomas},
booktitle={IEEE Winter Conference on Applications of Computer Vision},
year={2016}
}
Links and license
- Lab dataset page
- Dataset project and feature documentation
- Dataset license: Creative Commons Attribution 4.0. Code has its own license in the linked repository.
- Downloads last month
- 8,023