Skip to content

Latest commit

 

History

History
93 lines (73 loc) · 3.65 KB

DATASETS.md

File metadata and controls

93 lines (73 loc) · 3.65 KB
Raw

Data Preparation

To prepare the datasets, we prefer to organize the dataset as follows:

  • Standard Folder. Put all videos in the videos folder, and prepare the annotation files as train.txt and val.txt. Please make sure the folder looks like this: 
    $ ls /PATH/TO/videos | head -n 2
a.mp4
b.mp4

$ head -n 2 /PATH/TO/train.txt
a.mp4 0
b.mp4 2

$ head -n 2 /PATH/TO/val.txt
c.mp4 1
d.mp4 2

Additionally, we provide prepared annotation text files containing train/test splits for all datasets in datasets_splits folder.

The instructions to prepare each dataset are detailed below.

HMDB51

  • Download and unrar the HMDB51 dataset from the official website using this link.
  • Arrange the videos of all classes in a single folder structure as shown above.
  • Directly use the annotation files provided in datasets_splits/hmdb_splits folder.

UCF101

  • Download and unrar UCF101 dataset from the official website using this link
  • After extracting, the folder will be already in the required format (no need to arrange videos in a single folder).
  • Directly use the annotation files provided in datasets_splits/ucf_splits folder.

Kinetics

  • For downloading and preparing K400 and K600 datasets, we follow the steps provided in CVDF.
  • Run the below commands:
git clone https:/cvdfoundation/kinetics-dataset.git
cd kinetics-dataset
# Download k400
bash ./k400_downloader.sh
bash ./k400_extractor.sh
# Download k600
bash ./k600_downloader.sh
bash ./k600_extractor.sh
  • All videos of training set should be present in train and replacement folder and videos for validation set should be present in val folder.
  • replacement is a set of replacement videos for corrupted ones in K400 train set. After downloading and extracting k400 datasets from CVDF, move replacement_for_corrupted under replacement folder.
  • Overall, the directory structure should look like:
k400/
|–– train/ # contains videos for training set
|   |–– ZzZfJghfIz8_000068_000078.mp4
|   |–– ZzzLmGNirTg_000001_000011.mp4
|   └–– ...
|–– val/ # contains videos for val set
|   |–– -kahgmRD-4g_000004_000014.mp4
|   |–– KaMfzeeYwWw_000042_000052.mp4
|   └–– ...
└–– replacement/ # contains replacement videos for training set
    └–– replacement_for_corrupted
        |–– 2lujXIXQIg0_000052_000062.mp4
        └–– ...
  • Directly use the annotation files provided in datasets_splits/k400_splits folder.

Something Something v2 (SSv2)

  • Download SSv2 dataset from the official website using this link
  • Arrange the videos of all classes in a single folder structure similar to HMDB51 dataset.
  • Directly use the annotation files provided in datasets_splits/ssv2_splits

NOTE: The dataloader directly loads videos in an online fashion using decord.

TC-CLIP and its variants uses text-label embeddings for action classification. All labels files have been prepared and provided in labels/ folder. In order to create label file for your custom dataset, the format should look like this:

$ head -n 5 labels/kinetics_400_labels.csv
id,name
0,abseiling
1,air drumming
2,answering questions
3,applauding

The id indicates the class id, while the name denotes the text description.

Acknowledgements: ViFi-CLIP's repository.