This repository is to extract the dense frames of VISOR interpolations. VISOR website: (https://epic-kitchens.github.io/VISOR)
Before using the interpolations, you must run a correction script to to able to get the same JPG extracted frames. Please follow these steps:
- Download these 2 sparse jsons [ P02_01 , P03_14 ] to the script's directory.
- Run the correction script (
correct_interpolations.py) for each split (i.e train and val), the script has the following arguments:input_dir: a directory where the interpolations are, the interpolation json files must be directly under this directory.output_dir:a directory where the corrected interpolations would be stored, this could be the sameinput_dirdirectory.
Sample run:
For the train set interpolations:
python correct_interpolations.py --input_dir VISOR/Interpolations-DenseAnnotations/train --output_dir VISOR/Interpolations-DenseAnnotations/train
For the val set interpolations:
python correct_interpolations.py --input_dir VISOR/Interpolations-DenseAnnotations/val --output_dir VISOR/Interpolations-DenseAnnotations/val
Once that is done, the interpolations are corrcted and ready to be used.
Open generate_dense_frames.py script and edit the paths with your local paths. The script will contains generate_dense_images_from_video_jsons function with those arguments:
json_files_path: a path to JSON files you want to generate dense JPG images for.
The path file structure would be like:
-- /path/to/json/files/folder
---- P01_01_interpolations.json
---- P01_102_interpolations.json
---- PXX_XXX_interpolations.json
output_directory: a path where the output images would be generated.
sparse_rgb_images_root : a path to VISOR images. The sparse images would be copied from sparse_rgb_images_root to make sure that they're aliged with the sparse masks
The path file structure would be like:
-- /path/to/rbg_frames/folder
---- P01_01
------ P01_01_frame_0000xxxxxx.jpg
---- PXX_XXX
------ PXX_XXX_frame_0000xxxxxx.jpg
videos_path: a path to EPIC-KITCHENS videos. Videos (bulk or one by one) could be downloaded from there: EPIC-KITCHENS-100, or from there (code): EPIC-KITCHENS-100.
The path file structure would be like:
-- /path/to/videos/folder
---- P01_01.MP4
---- P01_102.MP4
---- PXX_XX(X).MP4
output_resolution (optional): output resolution of the generated images. Default is (854x480) which is the interpolation resolution
After identifying all arguments, just run python generate_dense_frames.py and you should get the dense images of the choosen videos. The images should be found in your output_directory.
In case you're interested in the timestamps of VISOR frames (as the original EPIC-KITCHENS videos have different frame rate), frames_to_timestamps.json contains the frame-to-timestamp mappings for train/val/test.
Ahmad Darkhalil* (1)
Dandan Shan* (2)
Bin Zhu* (1)
Jian Ma* (1)
Amlan Kar (3)
Richard Higgins (2)
Sanja Fidler (3)
David Fouhey (2)
Dima Damen (1)
1: University of Bristol, United Kingdom
2: University of Michigan, United States
3: University of Toronto, Canada
When using VISOR annotations, kindly reference both the annotations paper (VISOR) and the EPIC-KITCHENS-100 paper as follows:
-
Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, Dima Damen (2022). EPIC-KITCHENS VISOR Benchmark VIdeo Segmentations and Object Relations. Early Access OpenReview.
-
Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, Michael Wray (2022). Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100. International Journal of Computer Vision (IJCV) vol 130, pp 33-55.
All annotations are owned by the University of Bristol under the research license agreement 2021 - 3107, signed by all parties on Jan 2022.
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.
For commercial licenses, contact the University of Bristol at: uob-epic-kitchens@bristol.ac.uk
This deposit contains manually collected pixel-level segmentations for a subset of videos from the dataset EPIC-KITCHENS-100. These represent 30 recorded hours (train/val) and 6 hours of test (only images released).
This README contains information about annotations format. Please see epic-kitchens VISOR website for additional details, open challenges and the latest starter code.
The original videos can be downloaded from two separate DOIs: http://dx.doi.org/10.5523/bris.3h91syskeag572hl6tvuovwv4d and http://dx.doi.org/10.5523/bris.2g1n6qdydwa9u22shpxqzp0t8m
However, we also release sparse annotation frames for all splits of VISOR (train/val/test) with the following structure:
sparse
-- train
---- P01
-------P01_01.zip
-- val
-- test
We provide a mapping from these frames to the originally released rgb_frames in EPIC-KITCHENS. This avoids any differences in the frame extraction if any.
As explained in Sec 2.1, we provide manually annotated sparse segmentations, at the rate of (roughly) 2 frames per action clip. We use JSON formatting annotations, one per video, with the following fields
image
-- image_path: the folder and file name in the VISOR file structure. Note that this might not exactly match the EPIC-KITCHENS frame numbers (see Sec frame extraction for correspondences)
-- name: image filename
-- subsequence: Refer to Sec 2.1 where we define subsequences within videos with consistent set of entities. This field refers to the sequence number and is used for the VOS benchmark
-- video: video filename - matches EPIC-KITCHENS-100 video file name
annotations [one per mask]
-- id: Unique mask ID
-- name: open vocabulary entity name
-- class_id: key for EPIC-KITCHENS class number reflecting closed vocabulary entity name (see Sec classes for correspondences)
-- exhaustive: flag (y/n) to indicate whether all instances of the entity have been segmented. When 'n' is selected, background non-active instances of the same entity are present. We use 'inconclusive' when a consensus amongst annotators could not be found
-- in_contact_object: [only for hand and on-hand gloves] this field offers one of several options: the ID of the object that the hand is in contact with in this frame, or hand-not-in-contact when the hand is not touching any mask, none-of-the-above when the hand is touching an object that is not segmented, or inconclusive when a consensus amongst annotators could not be found.
-- on_which_hand: [only for gloves] indicating whether the glove is worn on a hand and which hand side is it.
-- segments: polygon-based representation of each mask
As explained in Sec 2.4, two consecutive ground truth segmentations are used to interpolate intermediate masks. The interpolations are filtered and only high J&F scored interpolations are provided.
In addition to the flags above, additional flags are available in interpolations:
-- type: 1: start/end ground-truth frames, that have been filtered to only include entities that are present at both the start and end frames. These are replicated here. 0: interpolated frames
-- interpolation: unique incremental ID for each interpolation in the dataset
Note that we had to re-extract frames from the dataset due to a change in ffmpeg versions. We provide a file that maps frame numbers between VISOR and publicly released rgb_frames in EPIC-KITCHENS-55 and EPIC-KITCHENS-100. You can find this at: frame_mapping.json
As new open-vocabulary entities have been added with the VISOR segmentations, we provide an updated noun classes for EPIC-KITCHENS. These include all new entities, mapped to the EPIC-KITCHENS-100 classes. We also add 4 additional classes: left hand, right hand, left glove, right glove - as prior classes didn't separate hand or glove sides. Note that we do not remove the original hand and glove classes for consistency with other benchmarks.