README.md

metadata

language:
  - en
license: cc-by-nc-sa-4.0
size_categories:
  - 10B<n<100B
task_categories:
  - other
pretty_name: ' Video2Reaction '
tags:
  - video
  - audio
  - text
dataset_info:
  features:
    - name: video_id
      dtype: string
    - name: reaction_dominant
      dtype: string
    - name: num_key_frames
      dtype: int32
    - name: clip_description_embedding
      sequence: float64
      length: 768
    - name: reaction_distribution
      sequence: float64
      length: 21
    - name: movie_genre
      sequence: float64
      length: 23
    - name: visual_feature
      dtype:
        array2_d:
          shape:
            - 176
            - 768
          dtype: float64
    - name: audio_acoustic_feature
      dtype:
        array2_d:
          shape:
            - 176
            - 1024
          dtype: float64
    - name: audio_semantic_feature
      dtype:
        array2_d:
          shape:
            - 176
            - 1024
          dtype: float64
  splits:
    - name: train
      num_bytes: 28780644620
      num_examples: 7243
    - name: val
      num_bytes: 4112655972
      num_examples: 1035
    - name: test
      num_bytes: 8225311923
      num_examples: 2070
  download_size: 8946422642
  dataset_size: 41118612515
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
      - split: test
        path: data/test-*

Video2Reaction

Data Structure

• /data folder uploaded using push_to_hub python function. It is a duplicated version of the following content in hf autogenerated parquet file format

• Same data of the original file format are uploaded to the root directory

  • {split}_vit_bert-base-uncased_clap_general_hubert_large.pt is a python dict that saves the torch tensor of the following latent features:
    • visual_feature
    • audio_acoustic_feature
    • audio_semantic_feature
    • clip_description_embedding
  • The first three features' first dim is the time dim and it is not a fixed size dim, and it can range from 16 to 176. In the parquet version to fit into huggingface auto generating system (to get croissant meta) we pad every feature's time dim to the max length (176)
  • {split}.json is the original meta file recording the video information

• Code repo: https://github.com/wm-bit/video2reaction

• REACTION_CLASSES

REACTION_CLASSES = ['sadness', 'disgust', 'grief', 'fear', 'disapproval', 'disappointment', 'embarrassment', 'nervousness', 'annoyance', 'anger', 'confusion', 'realization', 'caring', 'curiosity', 'relief', 'approval', 'surprise', 'excitement', 'amusement', 'admiration', 'joy']

• MOVIW_GENRES

MOVIW_GENRES = ['Music', 'Family', 'Crime', 'Thriller', 'Action', 'Western', 'Sci-Fi', 'Short', 'History', 'Adventure', 'Fantasy', 'Romance', 'Film-Noir', 'Biography', 'Comedy', 'Musical', 'War', 'Horror', 'Animation', 'Documentary', 'Sport', 'Mystery', 'Drama']

• SENTIMENT_2_FINER_GRAINED_MAPPING

SENTIMENT_2_FINER_GRAINED_MAPPING = {
"positive": ["amusement", "excitement", "joy", "love", "desire", "optimism", "caring", "pride", "admiration", "gratitude", "relief", "approval"],
"negative": ["fear", "nervousness", "remorse", "embarrassment", "disappointment", "sadness", "grief", "disgust", "anger", "annoyance", "disapproval"],
"ambiguous": ["realization", "surprise", "curiosity", "confusion"]
}