EPIC KITCHENS-100 Dataset

EPIC-KITCHENS-100 is the largest dataset in first-person (egocentric) vision; itself an extension of the EPIC-KITCHENS-55 dataset (formally known as EPIC-KITCHENS-2018).

Authors

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

(1 University of Bristol)
(2 University of Catania)

Contact: [email protected]

Citing

When using the dataset, kindly reference:

@article{damen2022rescaling,
   title={Rescaling Egocentric Vision},
   author={Damen, Dima and Doughty, Hazel and Farinella, Giovanni Maria  and and Furnari, Antonino 
           and Ma, Jian and Kazakos, Evangelos and Moltisanti, Davide and Munro, Jonathan 
           and Perrett, Toby and Price, Will and Wray, Michael},
           journal={International Journal of Computer Vision},
           volume={130},
           number={1},
           pages={33--55},
           year={2022},
           publisher={Springer}
}

Erratum

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, we below detail how you can fix the error at your end, until we publish replacement frames for downloading.

Download the videos P01_109.MP4 and P27_103.MP4. Then set up a directory like so:

$ mkdir -p rgb/{P01_109,P27_103}
$ mkdir -p flow/{P01_109,P27_103}
$ mkdir videos
$ mv /path/to/{P01_109,P27_103}.MP4 videos

You will need docker setup on your machine to extract the frames and flow.

RGB

$ docker run --gpus "device=0" \
     -it \
     --rm \
     -v "$PWD:/workspace" \
     willprice/nvidia-ffmpeg \
     -hwaccel cuvid \
     -c:v hevc_cuvid \
     -i /workspace/videos/P27_103.MP4 \
     -vf 'scale_npp=-2:256:interp_algo=super,hwdownload,format=nv12' \
     -qscale:v 4 \
     -r 50 /workspace/rgb/P27_103/frame_%010d.jpg

$ docker run --gpus "device=0" \
     -it \
     --rm \
     -v "$PWD:/workspace" \
     willprice/nvidia-ffmpeg \
     -hwaccel cuvid \
     -c:v hevc_cuvid \
     -i /workspace/videos/P01_109.MP4 \
     -vf 'scale_npp=-2:256:interp_algo=super,hwdownload,format=nv12' \
     -qscale:v 4 \
     -r 50 /workspace/rgb/P01_109/frame_%010d.jpg

Flow

$ docker run --gpus "device=0" \
     -it \
     --rm \
     -v "$PWD/rgb/P01_109:/input" \
     -v "$PWD/flow/P01_109:/output" \
     willprice/furnari-flow \
     frame_%010d.jpg -g 0 -s 1 -d 1 -b 8

$ docker run --gpus "device=0" \
     -it \
     --rm \
     -v "$PWD/rgb/P27_103:/input" \
     -v "$PWD/flow/P27_103:/output" \
     willprice/furnari-flow \
     frame_%010d.jpg -g 0 -s 1 -d 1 -b 8

Dataset Details

The EPIC-KITCHENS-100 dataset is an extension of the EPIC-KITCHENS-55 dataset. Videos are distinguished as follows:

PXX_YY.MP4 videos originate from EPIC-KITCHENS-55.
PXX_1YY.MP4 videos originate from the extension collected for EPIC-KITCHENS-100 (thus represent new videos).

The dataset currently has 6 active benchmarks:

Action Recognition
Weakly Supervised Action Recognition
Action Detection
Action Anticipation
Unsupervised Domain Adaptation
Multi-Instance Retrieval

We provide csv files for the train/val/test sets of each benchmark detailed below for ease of use, see Important Files for more information.

Ground truth is provided for action segments as action/verb/noun labels along with the start and end times of the segment.

We also provide automatic annotations in the form of object masks and hand/object BBoxes. See automatic annotations for more details.

back to top

Quick Start

Here you can download the annotation files for all of the challenges. For more information on each challenge, please see the paper here. A download script is provided for the videos, RGB Frames and Flow frames here.

Action Recognition Challenge

Download the videos/RGB/Flow frames here with the following command:

python epic_downloader.py --videos --rgb-frames --flow-frames

Download the Action Recognition train/val/test files.
Enjoy the EPIC-KITCHENS-100 dataset in your favourite action recognition model, see the paper for details on the models we used for this baseline. Models trained on EPIC-KITCHENS-55 can be found here as a starting point.

Weakly Supervised Action Recognition Challenge

Download the videos/RGB/Flow frames here with the following command:

python epic_downloader.py --videos --rgb-frames --flow-frames

This challenge uses the Action Recognition files, download the train/val/test files.
The weakly supervised challenge uses the narration timestamp, not the the start/end times of the action. Therefore a simple baseline would be to modify an action recognition model to use the surrounding 5s worth of frames. See the paper for details on the models we used for this baseline.

Action Detection Challenge

Download the videos/RGB/Flow frames here with the following command:

python epic_downloader.py --videos --rgb-frames --flow-frames

This challenge uses the Action Recognition files, download the train/val/test files.
Train an action proposal network on the EPIC-KITCHENS-100 train set, for example this model. This model predicts action-agnostic segments which still need to be classified.
Use your favourite action recognition model to classify the proposals (example models).

Action Anticipation Challenge

Download the videos/RGB/Flow frames here with the following command:

python epic_downloader.py --videos --rgb-frames --flow-frames

This challenge uses the Action Recognition files, download the train/val/test files.
A simple baseline for this task is to train an action recognition model (example models here) on the 5 seconds that precede an action with a 1 second gap. For example, an action that starts at 20.00s in a video would see frames between 14.00s and 19.00s.

Unsupervised Domain Adaptation Challenge

The unsupervised domain adaptation challenge tests how models can cope with similar data collected 2 years later on the task of action recognition.

Download the videos/RGB/Flow frames here with the following command:

python epic_downloader.py --videos --rgb-frames --flow-frames --domain-adaptation

Download the Unsupervised Domain Adaptation source train/target train/source_test/target test/source val/target val files.
Extract video features (for all six splits) using an off-the-shelf model trained on EPIC-KITCHENS-55 (example model).
A simple baseline is using a domain discriminator (prediciting whether a video came from the source, EPIC-KITCHENS-55, or the target, EPIC-KITCHENS-100) to align the two domains. See the paper for details on the models we used for this baseline.

IMPORTANT NOTE ON HYPER-PARAMETER TUNING. As the target domain is unlabelled, the training splits cannot be used for hyper-parameter tuning. You must use the validation splits to choose hyper-parameters. The procedure for hyper-parameter tuning and training is as follows:

Train your model on source val with unlabelled data from target val.
Evaluate your model on target val using the labels provided (these labels should not be used during training).
Select hyper-parameters based on the performance on target val.
Re-train your model on source train/target train with selected hyper-parameters.
Evaluate the re-trained model on target test to produce action predictions for the challenge leaderboard.

It is optional but highly ecouraged to evalute the performance on source_test to compare source domain performances.

Multi-Instance Retrieval Challenge

NOTE 30/09/2020 There was an error in the creation of the sentence files for the retrieval challenge. Please download the new sentence dataframes.

Download the videos/RGB/Flow frames here with the following command:

python epic_downloader.py --videos --rgb-frames --flow-frames --action-retrieval

Download the Multi-Instance Retrieval train/test files.
Extract video features (for both the train and test set) using an off-the-shelf model trained on EPIC-KITCHENS-55 (example model).
Extract word2vec features for the captions from both the train and test set (example models).
Enjoy the EPIC-KITCHENS-100 dataset in your favourite video retrieval model, see the paper for details on the models we used for this baseline.

back to top

Important Files

For ease of use, download scripts are provided to download the videos and RGB/Flow frames. (see file downloads for more details). We direct the reader to RDSF for the full release of videos and RGB/Flow frames. We provide html and pdf alternatives to this README which are auto-generated.

README.md (this file)
README.pdf
README.html
license.txt
EPIC_100_train.csv (info) (Pickle)
EPIC_100_validation.csv (info) (Pickle)
EPIC_100_test_timestamps.csv (info) (Pickle)
EPIC_100_noun_classes.csv (info)
EPIC_100_verb_classes.csv (info)

Additional Files

back to top

File Structure

EPIC_100_train.csv

This CSV file contains the action annotations for the training set and contains 15 columns:

Column Name	Type	Example	Description
`narration_id`	string	`P01_01_0`	Unique ID for the segment as a string with participant ID and video ID.
`participant_id`	int	`P01`	ID of the participant (unique per participant).
`video_id`	string	`P01_01`	ID of the video where the segment originated from (unique per video).
`narration_timestamp`	string	`00:00:01.089`	Timestamp of when the original narration was recorded in `HH:mm:ss.SSS`.
`start_timestamp`	string	`00:00:00.14`	Start time in `HH:mm:ss.SS` of the action segment.
`stop_timestamp`	string	`00:00:03.37`	End time in `HH:mm:ss.SS` of the action segment.
`start_frame`	int	`8`	Start frame of the action.
`stop_frame`	int	`202`	End frame of the action.
`narration`	string	`open door`	Transcribed description of the English narration provided by the participant.
`verb`	string	`open`	Parsed verb from the narration.
`verb_class`	int	`3`	Numeric ID of the verb's class.
`noun`	string	`door`	First parsed noun from the narration.
`noun_class`	int	`3`	Numeric ID of the first noun's class.
`all_nouns`	list of string (1 or more)	`[door]`	List of all parsed nouns within the narration.
`all_noun_classes`	list of int (1 or more)	`[3]`	Numeric ID of all of the parsed noun's classes.

Back to Important Files

EPIC_100_validation.csv

This CSV file contains the action annotations for the validation set and contains 15 columns:

Column Name	Type	Example	Description
`narration_id`	string	`P01_01_11`	Unique ID for the segment as a string with participant ID and video ID.
`participant_id`	int	`P01`	ID of the participant (unique per participant).
`video_id`	string	`P01_11`	ID of the video where the segment originated from (unique per video).
`narration_timestamp`	string	`00:00:00.560`	Timestamp of when the original narration was recorded in `HH:mm:ss.SSS`.
`start_timestamp`	string	`00:00:00.00`	Start time in `HH:mm:ss.SS` of the action segment.
`stop_timestamp`	string	`00:00:01.89`	End time in `HH:mm:ss.SS` of the action segment.
`start_frame`	int	`1`	Start frame of the action.
`stop_frame`	int	`113`	End frame of the action.
`narration`	string	`take plate`	Transcribed description of the English narration provided by the participant.
`verb`	string	`take`	Parsed verb from the narration.
`verb_class`	int	`0`	Numeric ID of the verb's class.
`noun`	string	`plate`	First parsed noun from the narration.
`noun_class`	int	`2`	Numeric ID of the first noun's class.
`all_nouns`	list of string (1 or more)	`[plate]`	List of all parsed nouns within the narration.
`all_noun_classes`	list of int (1 or more)	`[2]`	Numeric ID of all of the parsed noun's classes.

Back to Important Files

EPIC_100_test_timestamps.csv

This CSV file contains the action annotations for the testing set and contains 9 columns:

Column Name	Type	Example	Description
`narration_id`	string	`P01_101_0`	Unique ID for the segment as a string with participant ID and video ID.
`participant_id`	int	`P01`	ID of the participant (unique per participant).
`video_id`	string	`P01_101`	ID of the video where the segment originated from (unique per video).
`narration_timestamp`	string	`00:00:02.851`	Timestamp of when the original narration was recorded in `HH:mm:ss.SSS`.
`start_timestamp`	string	`00:00:02.86`	Start time in `HH:mm:ss.SSS` of the action segment.
`stop_timestamp`	string	`00:00:03.87`	End time in `HH:mm:ss.SSS` of the action segment.
`start_frame`	int	`143`	Start frame of the action.
`stop_frame`	int	`193`	End frame of the action.

Back to Important Files

EPIC_100_noun_classes.csv

This CSV file contains information on the 300 noun classes and contains 4 columns.

Column Name	Type	Example	Description
`id`	int	`222`	Unique ID for the noun class.
`key`	string	`label`	Key used for the noun class (all keys are a member of their own class).
`instances`	list of string (1 or more)	`"['label', 'sticker']"`	All nouns within the class, including the key.
`category`	string	`materials`	Name of the higher-level noun category that this noun class belongs to.

Back to Important Files

EPIC_100_verb_classes.csv

This CSV file contains information on the 97 verb classes and contains 4 columns.

Column Name	Type	Example	Description
`id`	int	`79`	Unique ID for the verb class.
`key`	string	`let-go`	Key used for the verb class (all keys are a member of their own class).
`instances`	list of string (1 or more)	`"['let', 'let-go']"`	All verbs within the class, including the key.
`category`	string	`leave`	Name of the higher-level verb category that this verb class belongs to.

Back to Important Files

EPIC_100_uda_source_train.csv

This CSV file contains the action annotations for the source training set used for Unsupervised Domain Adaptation and contains 15 columns:

Column Name	Type	Example	Description
`narration_id`	string	`P01_01_0`	Unique ID for the segment as a string with participant ID and video ID.
`participant_id`	int	`P01`	ID of the participant (unique per participant).
`video_id`	string	`P01_01`	ID of the video where the segment originated from (unique per video).
`narration_timestamp`	string	`00:00:01.089`	Timestamp of when the original narration was recorded in `HH:mm:ss.SSS`.
`start_timestamp`	string	`00:00:00.14`	Start time in `HH:mm:ss.SS` of the action segment.
`stop_timestamp`	string	`00:00:03.37`	End time in `HH:mm:ss.SS` of the action segment.
`start_frame`	int	`8`	Start frame of the action.
`stop_frame`	int	`202`	End frame of the action.
`narration`	string	`open door`	Transcribed description of the English narration provided by the participant.
`verb`	string	`open`	Parsed verb from the narration.
`verb_class`	int	`3`	Numeric ID of the verb's class.
`noun`	string	`door`	First parsed noun from the narration.
`noun_class`	int	`3`	Numeric ID of the first noun's class.
`all_nouns`	list of string (1 or more)	`[door]`	List of all parsed nouns within the narration.
`all_noun_classes`	list of int (1 or more)	`[3]`	Numeric ID of all of the parsed noun's classes.

Note that this file contains only videos from EPIC-KITCHENS-55 which is used as the source domain.