Our dataset is built upon the Something-Something V2 and COIN dataset

Please download the repectives dataset from the original works:

• COIN (https://coin-dataset.github.io/)
• SSv2 (https://www.qualcomm.com/developer/software/something-something-v-2-dataset/downloads)

We provide files for Effect2M splits and a annotation file for segmentation on coin dataset.

After downloading COIN dataset, use the COIN_effect.csv file to segment the videos to clips of actually action occurence.

From COIN_effect.csv use the following columns:
vid_id - Name to save the segmented video clips
ori_video - original COIN video name
start - start frame of segment
end - end frame of segment

Segment the videos into clips following the start and end , saving the new clips with the vid_id as the filename.

We split the training annotation file into 3 csv due to the size.

Annotations consists of the following columns:
Class - action label
State - ID of video used for the pre and post action states
Positive - ID of video used for the Postive action sample for the action
Augs - To updated
Negative1-3 - ID of video used for the Negative action samples for the action
Neg_cls1-3 - action label for video used for the Negative action samples for the action
dataset - Original dataset which video sample belong to

This repo is built based on the fantastic works of SSv2, COIN. We thank the authors for their great work and open-sourcing their codebase.