EPIC-KITCHENS-55 is the largest dataset in first-person (egocentric) vision; 55 hours of multi-faceted, non-scripted recordings in native environments - i.e. the wearers' homes, capturing all daily activities in the kitchen over multiple days. Annotations are collected using a novel `live' audio commentary approach.

Dima Damen (1) Hazel Doughty (1) Giovanni Maria Farinella (3) Sanja Fidler (2) Antonino Furnari (3) Evangelos Kazakos (1) Davide Moltisanti (1) Jonathan Munro (1) Toby Perrett (1) Will Price (1) Michael Wray (1)

• (1 University of Bristol)
• (2 University of Toronto)
• (3 University of Catania)

Contact: [email protected]

When using the dataset, kindly reference:

@INPROCEEDINGS{Damen2018EPICKITCHENS,
   title={Scaling Egocentric Vision: The EPIC-KITCHENS Dataset},
   author={Damen, Dima and Doughty, Hazel and Farinella, Giovanni Maria  and Fidler, Sanja and 
           Furnari, Antonino and Kazakos, Evangelos and Moltisanti, Davide and Munro, Jonathan 
           and Perrett, Toby and Price, Will and Wray, Michael},
   booktitle={European Conference on Computer Vision (ECCV)},
   year={2018}
} 

(Check publication here)

We provide ground truth for action segments and object bounding boxes.

• Objects: Full bounding boxes of narrated objects for every annotated frame.
• Actions: Split into narrations and action labels:
  • Narrations containing the narrated sentence with the timestamp.
  • Action labels containing the verb and noun labels along with the start and end times of the segment.

The dataset is comprised of three splits with the corresponding ground truth:

• Training set - Full ground truth.
• Seen Kitchens (S1) Test set - Start/end times only.
• Unseen Kitchens (S2) Test set - Start/end times only.

Initially we are only releasing the full ground truth for the training set in order to run action and object challenges.

• README.md (this file)
• README.html
• README.pdf
• license.txt
• EPIC_train_action_labels.csv (Info) (Pickle)
• EPIC_train_object_labels.csv (Info)
• EPIC_test_s1_timestamps.csv (Info) (Pickle)
• EPIC_test_s2_timestamps.csv (Info) (Pickle)
• EPIC_train_object_action_correspondence.csv (Info) (Pickle)
• EPIC_test_s1_object_action_correspondence.csv (Info) (Pickle)
• EPIC_test_s2_object_action_correspondence.csv (Info) (Pickle)
• EPIC_test_s1_object_video_list.csv (Info)
• EPIC_test_s2_object_video_list.csv (Info)
• EPIC_noun_classes.csv (Info)
• EPIC_verb_classes.csv (Info)

• EPIC_train_invalid_labels.csv (Info) (Pickle)
• EPIC_train_action_narrations.csv (Info)
• EPIC_descriptions.csv (Info)
• EPIC_many_shot_verbs.csv (Info)
• EPIC_many_shot_nouns.csv (Info)
• EPIC_many_shot_actions.csv (Info)
• EPIC_video_info.csv (info)

We direct the reader to RDSF for the videos and rgb/flow frames.

We provide html and pdf alternatives to this README which are auto-generated.

CSV file containing 14 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file containing 14 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file containing 5 columns:

Note: The start/end timestamp refers to the start/end time of the narration, not the action itself.

CSV file containing 6 columns:

CSV file containing 5 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file containing 5 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file containing 5 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file listing the videos used to obtain the object s1 test frames. The frames can be obtained from RDSF under object_detection_images/test. Please test all frames from this folder for the videos listed in this csv.

CSV file listing the videos used to obtain the object s2 test frames. The frames can be obtained from RDSF under object_detection_images/test. Please test all frames from this folder for the videos listed in this csv.

CSV file containing 7 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file containing 7 columns:

Please note we have included a python pickle file for ease of use. This includes a pandas dataframe with the same layout as above. This pickle file was created with pickle protocol 2 on pandas version 0.22.0.

CSV file containing 3 columns:

Note: a colon represents a compound noun with the more generic noun first. So pan:dust should be read as dust pan.

CSV file containing 3 columns:

CSV file containing 4 columns:

CSV file containing the many shot verbs. A verb class is considered many shot if it appears more than 100 times in training. (NOTE: this file is derived from EPIC_train_action_labels.csv, checkout the accompanying notebook demonstrating how we compute these classes)

CSV file containing the many shot nouns. A noun class is considered many shot if it appears more than 100 times in training. (NOTE: this file is derived from EPIC_train_action_labels.csv, checkout the accompanying notebook demonstrating how we compute these classes)

CSV file containing the many shot actions. An action class (composed of a verb class and noun class) is considered many shot if BOTH the verb class and noun class are many shot AND the action class appears in training at least once. (NOTE: this file is derived from EPIC_train_action_labels.csv, checkout the accompanying notebook demonstrating how we compute these classes)

CSV file containing information for each video

Due to the size of the dataset we provide scripts for downloading parts of the dataset:

• videos (750GB)
• frames (320GB)
  • rgb-frames (220GB)
  • flow-frames (100GB)
• object annotation images (80GB)

Note: These scripts will work for Linux and Mac. For Windows users a bash installation should work.

These scripts replicate the folder structure of the dataset release, found here.

If you wish to download part of the dataset instructions can be found here.

Videos are recorded in 1080p at 59.94 FPS on a GoPro Hero 5 with linear field of view. There are a minority of videos which were shot at different resolutions, field of views, or FPS due to participant error or camera. These videos identified using ffprobe are:

• 1280x720: P12_01, P12_02, P12_03, P12_04.
• 2560x1440: P12_05, P12_06
• 29.97 FPS: P09_07, P09_08, P10_01, P10_04, P11_01, P18_02, P18_03
• 48 FPS: P17_01, P17_02, P17_03, P17_04
• 90 FPS: P18_09

The GoPro Hero 5 was also set to drop the framerate in low light conditions to preserve exposure leading to variable FPS in some videos. If you wish to extract frames we suggest you resample at 60 FPS to mitigate issues with variable FPS, this can be achieved in a single step with FFmpeg:

ffmpeg -i "P##_**.MP4" -vf "scale=-2:256" -q:v 4 -r 60 "P##_**/frame_%010d.jpg"

where ## is the Participant ID and ** is the video ID.

Optical flow was extracted using a fork of gpu_flow made available on github. We set the parameters: stride = 2, dilation = 3, bound = 25 and size = 256.

All files in this dataset are copyright by us and published under the Creative Commons Attribution-NonCommerial 4.0 International License, found here. This means that you must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use. You may not use the material for commercial purposes.

EPIC-KITCHENS-55 and EPIC-KITCHENS-100 were collected as a tool for research in computer vision, however, it is worth noting that the dataset may have unintended biases (including those of a societal, gender or racial nature).

See release history for changelog.