We need to provide support for users to compute metrics for the EPIC-Kitchens 100 challenges.

Is there any example code to gulp dataset with manual downloading.
For example, i have downloaded all the tar files for rgb and flow.
I have also extracted them into frames. But I don't know how to gulp it.

The frame iterators in each modality only go up to the stop_frame - 1, these should probably be inclusive bounds i.e. range(start_frame, stop_frame - 1).

https://github.com/epic-kitchens/epic-lib/blob/master/epic_kitchens/video.py#L41
https://github.com/epic-kitchens/epic-lib/blob/master/epic_kitchens/video.py#L54

Spotted by @ekazakos

Provide methods to compute:

• Top-1 accuracy
  • verbs
  • nouns
  • actions
• Top-5 accuracy
  • verbs
  • nouns
  • actions
• Per-class recall on many shot classes
  • verbs
  • nouns
  • actions
• Per-class precision on many shot classes
  • verbs
  • nouns
  • actions

Given a groundtruth data frame containing the correct verb, noun, and action for a given action segment (identified by UID)

How did you calculate the top-5 accuracy for "action" which is a pair of (noun, verb)?

Such that both EK-55 and EK-100 can be used throughout the lib

When importing epic_kitchens.gulp.visualisation, I get an error message :
imageio.ffmpeg.download() has been deprecated. Use 'pip install imageio-ffmpeg' instead.'
but installing this library does not help.

It looks that you are depending on an earlier version of imageio, like the 2.4.1
Maybe requirements.txt should be edited accordingly

In [30]: import epic_kitchens.gulp.visualisation
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-30-8a3f6f3f60a0> in <module>()
----> 1 import epic_kitchens.gulp.visualisation

/usr/local/lib/python3.5/dist-packages/epic_kitchens/gulp/visualisation.py in <module>()
      2 import numpy as np
      3 from abc import ABC
----> 4 from moviepy.editor import ImageSequenceClip
      5 from moviepy.video.io.html_tools import ipython_display
      6 from typing import List, Union, Tuple

/usr/local/lib/python3.5/dist-packages/moviepy/editor.py in <module>()
     24 # Checks to see if the user has set a place for their own version of ffmpeg
     25 if os.getenv('FFMPEG_BINARY', 'ffmpeg-imageio') == 'ffmpeg-imageio':
---> 26     imageio.plugins.ffmpeg.download()
     27 
     28 # Clips

/usr/local/lib/python3.5/dist-packages/imageio/plugins/ffmpeg.py in download(directory, force_download)
     38 def download(directory=None, force_download=False):  # pragma: no cover
     39     raise RuntimeError(
---> 40         "imageio.ffmpeg.download() has been deprecated. "
     41         "Use 'pip install imageio-ffmpeg' instead.'"
     42     )

RuntimeError: imageio.ffmpeg.download() has been deprecated. Use 'pip install imageio-ffmpeg' instead.'

My solution

to 'solve' this I installed the aforementioned version:
sudo pip3 install imageio==2.4.1

Hi,

Please correct me if im wrong, but currently there isnt any support for object detection experiments in the library right?

Hello,

Would it make sense to Gulp audio too ?
I'm doing it, but I wonder if there is a specific reason not to do it...

Or at least find a way to unit test it, although might not be worth it due to the complexities of filesystems.

small bug in epic_kitchens.time.flow_frame_count(start_frame, stride, dilation)

if start_frame == 1 it will return -1

Is there a particular reason for which the default data_dir isn't a hidden folder?

• It seems a bit intrusive to create a visible folder in the $HOME of the user. Other libraries and apps (PyTorch, Jupyter, etc.) use the hidden folder pattern.

• I do appreciate that we can set the data_dir.

BTW, thanks for providing this library to easily interact with the dataset.

Support downloading for the new EPIC-Kitchens 100 annotations

Hi,
I've use the following method to extract optical flow and RGB images:

$ python -m epic_kitchens.preprocessing.split_segments \
    P01 \
    root_dir/P01 \
    output_dir/P01 \
    EPIC_train_action_labels.pkl \
    flow \
    --fps 60 \
    --frame-format 'frame_%010d.jpg'

But it seems that optical flow images are extracted every two RGB images. Is it possible to extract optical flow image for every RGB image? Thank you!