We provide a python script to download both the EPIC KITCHENS-100 and EPIC KITCHENS-55 datasets, either in its entirety or in parts (e.g. only RGB frames). You can download data for a subset of participants, as well as a specific split for a certain challenge (e.g. the test split for action retrieval). You can also download only the dataset's extension, i.e. only the new data collected for EPIC KITCHENS-100.
The script requires Python 3.5+ and no external libraries.
We are aware of an issue occurring when using the download script depending on the region, e.g.
Could not download file from https://data.bris.ac.uk/datasets/3h91syskeag572hl6tvuovwv4d/frames_rgb_flow/rgb/train/P01/P01_01.tar
Error: [Errno 104] Connection reset by peerWhile we fix this problem, please use Academic Torrents if you face this issue. You can find EPIC-KITCHENS-55 here and EPIC-KITCHENS-100 here.
Important: We have recently detected an error in our pre-extracted RGB and Optical flow frames for two videos in our dataset. This does not affect the videos themselves or any of the annotations in this github.
However, if you've been using our pre-extracted frames or object/hand detections and masks, you can use the script to download the correct files:
python epic_downloader.py --errataThis will download only the affected files:
P01/rgb_frames/P01_109.tarP01/flow_frames/P01_109.tarP01/hand-objects/P01_109.pklP01/masks/P01_109.pklP27/rgb_frames/P27_103.tarP27/flow_frames/P27_103.tarP27/hand-objects/P27_103.pklP27/masks/P27_103.pkl
If you had downloaded these files previously, they will be overwritten.
As of 26/09/2020, regardless of the --errata argument, the script will always download the correct version of the
above files.
As of 02/12/2020, the correct object/hand detections and masks will also be downloaded regardless of the --errata argument.
The script accepts a number of arguments that allow you to specify what you want to download: by default the script will download everything to your home directory.
We provide videos, RGB/optical flow frames, GoPro's metadata (for the extension only) and object detection frames (for EPIC KITCHENS-55's videos only). You can also download the consent form templates.
If you want to download only one (or a subset) of the above, you can do so with the following self-explanatory arguments:
--videos--rgb-frames--flow-frames--object-detection-images--masks--metadata--consent-forms
If you want to download only videos, then:
python epic_downloader.py --videosNote that these arguments can be combined to download multiple things. For example:
python epic_downloader.py --rgb-frames --flow-framesWill download both RGB and optical flow frames.
For more information about objects masks (which you download with --masks), check out the dedicated repository.
You can use the argument --participants if you want to download data for only a subset of the participants. Participants can be specified with their numerical or string ID.
You can specify a single participant, e.g. --participants 1 or --participants P01 for participant P01, or a comma-separated list of them, e.g. --participants 1,2,3 or --participants P01,P02,P03 for participants P01, P02 and P03
This argument can also be combined with the aforementioned arguments. For example:
python epic_downloader.py --videos --participants 1,2,3Will download only videos from P01, P02 and P03.
You can use the argument --specific-videos if you want to download a specific subset of videos, without having to download all videos from certain participants. Videos must be specified with their string ID.
You can specifiy a single video e.g. --specific-videos P01_01 for video P01_01, or a comma-seperated list of them, e.g. --specific-videos P01_01,P02_122,P03_04 for videos P01_01, P02_122 and P03_04.
This argument can also be combined with the aforementioned arguments. For example:
python epic_downloader.py --videos --participants P01 --specific-videos P02_122,P03_04 Will download all the videos from participant P01, but only videos P02_122 and P03_04 from participants P02 and P03.
Note: if the same participant is specified in the partipant and specific-videos argument, the specific video of that participant will not be exclusively downloaded. Instead, all videos from that particpant will be downloaded e.g. --particpants P01 --specific-videos P01_01 will download all videos from participant P01, rather than just P01_01.
You can choose to download data for a specific split of one of the three challenges: action recognition, domain adaptation and action retrieval.
To specify a challenge, use the following arguments:
--action-recognition--domain-adaptation--action-retrieval
To specify a split, use one (or more) of the following arguments.
| argument | challenge(s) available |
|---|---|
--train |
action recognition, action retrieval |
--val |
action recognition only |
--test |
action recognition, action retrieval |
--source-train |
domain adaptation |
--source-test |
domain adaptation |
--target-train |
domain adaptation |
--target-test |
domain adaptation |
--val-source-train |
domain adaptation |
--val-source-test |
domain adaptation |
--val-target-train |
domain adaptation |
--val-target-test |
domain adaptation |
Combine challenge and split arguments to download the relevant splits. For example:
python epic_downloader.py --action-recognition --trainWill download all data for the training set split for action recognition.
python epic_downloader.py --action-retrieval --test --domain-adaptation --target-test Will download all data for the test set split for action retrieval, as well as all data for the target test split for domain adaptation.
Again, these arguments can be combined with all others. For example
python epic_downloader.py --flow-frames --participants 7 --action-recognition --testWill download only optical flow frames for P07's videos contained in the test split for action recognition.
Arguments can be passed in any order.
You can also specify only split arguments without a challenge. For example
python epic_downloader.py --test --videosWill download all videos in the test sets of the action recognition and action retrieval challenges.
python epic_downloader.py --train --valWill download all data for the training sets of action recognition and action retrieval, as well as the validation set for action recognition.
Refer to the table above to check what gets downloaded with each split argument.
These refer to unique splits which should be used only to validate hyper-parameters. For example, -- val-target-test will download the target test set used for validation. See the EPIC-KITCHENS-100 annotations repo here for more information about domain adaptation splits.
You can do so with the --extension-only and --epic55-only arguments. For example:
python epic_downloader.py --videos --extension-onlyWill download only the newly collected videos, while
python epic_downloader.py --participants 3 --epic55-onlyWill download all data from P03 collected for EPIC KITCHENS-55.
By default data will be downloaded to your home directory. You can change the output folder with the --output-path argument.
Data will be stored in a subdirectory named EPIC-KITCHENS.
Data will be split amongst participants under the output directory:
/home/username/EPIC-KITCHENS/
├── ConsentForms
└── P01
├── flow_frames
├── hand-objects
├── masks
├── meta_data
├── rgb_frames
└── videos
Each leaf directory will contain .tar, .mp4 or .pkl files accordingly.
Previously fully downloaded files will be skipped, so you can download large batches of files over multiple runs.
Errata files will be overwritten. Once you download the correct version of these files, they will be safely skipped in following runs.
Download speed might be (very) slow depending on the region.
In such case we recommend using Academic Torrents instead of this script. You can find EPIC-KITCHENS-55 here and EPIC-KITCHENS-100 here.