Number of Action Classes and Action Segments Slightly Different From the Paper.

[Guglielmo Camporese]

When I process the action segments from the CSVs (EPIC_100_train.csv and EPIC_100_validation.csv) I obtain slightly different numbers with respect to the EK-100 paper [arXiv].

I'm referring to:

• the number of train action segments,
• the number train action classes,
• the number validation action classes.

Here I reported the numbers from what I obtain and the paper.

Source	Partition	Action Segments	Action Classes
Paper [arXiv]	Train	67,219	3,550
Code below	Train	67,217	3,568

Source	Partition	Action Segments	Action Classes
Paper [arXiv]	Validation	9,668	1,344
Code below	Validation	9,668	1,352

Here the code for reproducing the numbers:

import pandas as pd

def get_actions(df):
    """Encode the action as '{v_class}_{n_class}'."""
    return [f'{v}_{n}' for v, n in zip(df['verb_class'], df['noun_class'])]

# Load CSVs and get actions
actions_train = get_actions(pd.read_csv('EPIC_100_train.csv'))
actions_validation = get_actions(pd.read_csv('EPIC_100_validation.csv'))

# Output
train_action_seg = len(actions_train) # -> 67217
train_action_classes = len(set(actions_train)) # -> 3568
validation_action_seg = len(actions_validation) # -> 9668
validation_action_classes = len(set(actions_validation)) # -> 1352

• Is this gap expected?
• What is the correct number of action classes?
• Could you share a csv where you identify the actions (a = [v_class, n_class])?

Best,

Guglielmo

[Michael Wray]

Hi Guglielmo,

Thanks for posting this, there is indeed a discrepancy between the numbers which stemmed from an internal pre-release version of the dataset which weren't updated for the latest arXiv release by mistake.

The numbers that you posted are indeed correct, i.e.

Training Set

# Action Segments: 67,217
# Action Classes: 3,568

Validation Set

# Action Classes: 1,352

Mike