Data Facts

June 9, 2021 ยท View on GitHub

Video Question Answering

TVQA

Paper, Dataset Website

TVQA is a video question answering dataset based on 6 TV shows. Its annotations can be downloaded by:

bash download_annotations/download_tv_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train122,039tvqa_train.jsonl
val15,253tvqa_val.jsonl
test15,253tvqa_test_release.jsonl

The files are in JSON Lines format. Each line can be loaded as a Python dict with the following entries:

{
  "a0": "Because Sheldon is being rude.",  # str, candidate answers
  "a1": "Because he doesn't like Sheldon.",
  "a2": "Because they are having an argument.",
  "a3": "Because Howard wanted to have a private meal with Raj.",
  "a4": "Because Sheldon won't loan him money for food.",
  "answer_idx": 2,  # int, ground-truth answer id
  "q": "Why is Howard ...",  # str, question
  "qid": 122039,  # int, unique question id
  "show_name": "The Big Bang Theory",  # name of the TV show
  "ts": "20.16-25.12",  # start and end timestamps of the question
  "vid_name": "s03e02_seg02_clip_10"  # video name
}

How2QA

Paper

How2QA is a video question answering dataset based on YouTube videos. Its annotations can be downloaded by:

bash download_annotations/download_how2_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train34,187how2qa_train_release.jsonl
val3,115how2qa_val_release.jsonl
test_public3,095how2qa_test_public_release.jsonl

The files are in JSON Lines format. Each line is formatted similarly to TVQA, which we detailed above.

VIOLIN

Paper

VIOLIN is a video-and-language inference dataset based on TV show and movie videos. Its annotations can be downloaded by:

bash download_annotations/download_violin_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train76,122violin_train_release.jsonl
val9,600violin_val_release.jsonl
test9,600violin_test_release.jsonl
test_private7,722violin_test_private_release.jsonl

The files are in JSON Lines format. Each line can be loaded as a Python dict:

{
  "paired_id": "dh_s08e07_clip_1374_1411-2-0",  # str, each true (false) statement is paired with a false (true) statement. The paired_id for it is dh_s08e07_clip_1374_1411-2-1. Note that paired_id is not provided in test_private set.
  "vid_name": "dh_s08e07_clip_1374_1411",  # str, video name
  "statement": "The woman with red hair ...",  # str, statement
  "example_id": 85354,  # int, unique example id
  "answer": 1  # int, ground-truth label, 1=true, 0=false
}

VLEP

Paper

VLEP is a video-and-language future event prediction dataset based on TV show and YouTube Vlog videos. Its annotations can be downloaded by:

bash download_annotations/download_vlep_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train20,142vlep_train_release.jsonl
dev4,392vlep_dev_release.jsonl
test4,192vlep_test_release.jsonl

The files are in JSON Lines format. Each line can be loaded as a Python dict:

{
  "example_id": 21411,  # int, unique example id
  "vid_name": "OO9kSxcT9Rg_subs_012_00:13:00_00:14:00_ep",  # video name.
  "ts": [10.54, 13.55],  # start and end time of the premise event (in seconds).
  "events": [  
    "The customer eats the food from the tray.",
    "The customer takes the tray of food."
  ],  # the two possible future events after the premise.
  "answer": 1,  # int, ground-truth answer idx, in {0, 1}
  "split": "dev"  # str, split
}

Video Retrieval

TVR

Paper, Dataset Website

TVR is a video moment retrieval dataset based on 6 TV shows (the same as TVQA). Its annotations can be downloaded by:

bash download_annotations/download_tv_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train87,175tvr_train_release.jsonl
val10,895tvr_val_release.jsonl
test15,253tvr_test_release.jsonl

The files are in JSON Lines format. Each line can be loaded as a Python dict with the following entries:

{
  "vid_name": "friends_s04e20_seg02_clip_07",  # str, video name
  "duration": 58.03,  # int, video duration
  "ts": [50.19, 57.73],  # list(int), start and end timestamps of the moment
  "desc": "Joey closes the magazine ...",  # str, text query
  "type": "v",  # str, query type, v=video, t=subtitle, vt=video+subtitle 
  "desc_id": 93717  # int, unique desc id
}

How2R

Paper

How2R is a video moment retrieval dataset based on YouTube videos. Its annotations can be downloaded by:

bash download_annotations/download_how2_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train27,135how2r_train_release.jsonl
val1,295how2r_val_1k_release.jsonl
test_public1,265how2r_test_public_1k_release.jsonl

The files are in JSON Lines format. Each line is formatted similarly to TVR, which we detailed above.

VATEX-EN-R

Paper, Dataset Website

VATEX-EN-R is a video retrieval dataset based on YouTube videos, originally created for video captioning. Its annotations can be downloaded by:

bash download_annotations/download_vatex_en_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train259,910vatex_en_r_train_release.jsonl
val30,000vatex_en_r_val_release.jsonl
test60,000vatex_en_r_test_public_release.jsonl

The files are in JSON Lines format. Each line is formatted similarly to TVR, which we detailed above.

YC2R

Paper, Dataset Website

YC2R is a video retrieval dataset based on YouTube cooking videos. Its annotations can be downloaded by:

bash download_annotations/download_yc2_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train10,337yc2r_train_release.jsonl
val3,492yc2r_val_release.jsonl
test1,604yc2r_test_release.jsonl

The files are in JSON Lines format. Each line is formatted similarly to TVR, which we detailed above.

video2dur_idx

Each of the retrieval datasets is also accompanied by a video2dur_idx file, containing mappings from video name to video duration and video idx (used for submission). For example, tvr_video2dur_idx.json is for TVR, each entry in this file is:

{
  "castle_s01e01_seg02_clip_20": [
    91.56,  # float, video duration 
    0  # video id, used for retrieval submission
  ],
  ...
}

Video Captioning

TVC

Paper, Dataset Website

TVC is a video captioning dataset based on 6 TV shows (the same as TVQA). It shares some of the annotations with TVR. Its annotations can be downloaded by:

bash download_annotations/download_tv_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train86,603tvc_train_release.jsonl
val10,895tvc_val_release.jsonl
test15,253tvc_test_release.jsonl

The files are in JSON Lines format. Each line can be loaded as a Python dict with the following entries:

{
  "vid_name": "grey_s03e12_seg02_clip_17",  # video id
  "duration": 94.03,  # float, duration of the video
  "ts": [35.73, 40.9],  # list(float), start and end timestamps of the clip
  "descs": [
    {
      "desc": "Meredith walks into the kitchen...", # str, description of the clip
      "desc_id": 62633,  # int, unique desc id
      "type": "v",  # str, desc type, v=video, t=subtitle, vt=video+subtitle 
      "from_retrieval": true  # bool, true if it is shared with TVR, false otherwise
    },
    ...  # each clip in train has 2 captions, each clip in {val, test} has 4 captions.
  ],
  "clip_id": 86602  # int, unique clip id
}

VATEX-EN-C

Paper, Dataset Website

VATEX-EN-C is a video captioning dataset based on YouTube videos (the same as VATEX-EN-R). Its annotations can be downloaded by:

bash download_annotations/download_vatex_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train25,991vatex_en_c_train_release.jsonl
val3,000vatex_en_c_val_release.jsonl
test_public6,000vatex_en_c_test_public_release.jsonl
test_private6,278vatex_en_c_test_private_release.jsonl

The files are in JSON Lines format. Each line is formatted similarly to TVC, which we detailed above.

YC2C

Paper, Dataset Website

YC2C is a video captioning dataset based on YouTube cooking videos (the same as YC2R). Its annotations can be downloaded by:

bash download_annotations/download_yc2_annotations.sh $DATA_DIR

$DATA_DIR is the directory storing the downloaded annotations. There are 3 annotation files as list below:

Split#examplesFilename
train25,991yc2c_train_release.jsonl
val3,000yc2c_val_release.jsonl
test6,278yc2c_test_release.jsonl

The files are in JSON Lines format. Each line is formatted similarly to TVC, which we detailed above.

Subtitle Files

For each video, we also provided their associated subtitle files. For example, tv_subtiles.jsonl contains the subtitle files for TVR, TVC, TVQA and part of VLEP. This file is also in JSON Lines format. Each line is:

{
  "vid_name": "house_s02e05_seg02_clip_11",  # str, video name
  "sub": [
    {
      "text": " Chase : That's all this is?",  # str, subtitle text
      "start": 0.862,  # float, start timestamp of the subtitle
      "end": 1.862  # float, end timestamp of the subtitle
    },
    ...
  ]
}

License

As per the original authors, the annotations for TVQA, TVR, TVC, VIOLIN, YouCookII, VLEP, How2QA, How2R are under CC BY-NC-SA 4.0 license, the annotations for VATEX are under CC BY 4.0.