dataset_info:
features:
- name: channel_id
dtype: string
- name: video_id
dtype: string
- name: segment_id
dtype: int64
- name: duration
dtype: string
- name: fps
dtype: int64
- name: conversation
list:
- name: end_time
dtype: float64
- name: speaker
dtype: int64
- name: start_time
dtype: float64
- name: text
dtype: string
- name: utterance_id
dtype: int64
- name: words
list:
- name: end_time
dtype: float64
- name: start_time
dtype: float64
- name: word
dtype: string
- name: facial_expression
list:
- name: features
sequence: float32
- name: frame
dtype: int64
- name: utt_id
dtype: int64
- name: body_language
list:
- name: features
sequence: float32
- name: frame
dtype: int64
- name: utt_id
dtype: int64
- name: harmful_utterance_id
sequence: int64
- name: speaker_bbox
list:
- name: bbox
sequence: int64
- name: frame_id
dtype: int64
splits:
- name: train
num_bytes: 144100517656
num_examples: 7985
- name: test
num_bytes: 31918682474
num_examples: 1993
download_size: 166967732474
dataset_size: 176019200130
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
Dataset Card for VENUS
Dataset Summary
Data from: Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues
@article{kim2025speaking,
title={Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues},
author={Kim, Youngmin and Chung, Jiwan and Kim, Jisoo and Lee, Sunghyun and Lee, Sangkyu and Kim, Junhyeok and Yang, Cheoljong and Yu, Youngjae},
journal={arXiv preprint arXiv:2506.00958},
year={2025}
}
We provide a multimodal large-scale video dataset based on nonverbal communication.
Please cite our work if you find our data helpful.
Our dataset collection pipeline and the model implementation that uses it are available at https://github.com/winston1214/nonverbal-conversation
Dataset Statistic
| Split | Channels | Videos | Segments (10 min) | Frames (Nonverbal annotations) | Utterances | Words |
|---|---|---|---|---|---|---|
| Train | ~ | ~ | ~ | ~ | ~ | |
| Test | ~ | ~ | ~ | ~ | ~ |
Language
English
Other Version
- VENUS-1K: This link
- VENUS-5K: This link
- VENUS-25K: This link
- VENUS-50K: This link (Comming Soon!)
- VENUS-100K (Full): This link (Comming Soon!)
Data Structure
Here's an overview of our dataset structure:
{
'channel_id': str, # YouTube channel ID
'video_id': str, # Video ID
'segment_id': int, # Segment ID within the video
'duration': str, # Total segment duration (e.g., '00:11:00 ~ 00:21:00')
'fps': int, # Frames per second
'conversation': [ # Conversation information (consisting of multiple utterances)
{
'utterance_id': int, # Utterance ID
'speaker': int, # Speaker ID (represented as an integer)
'text': str, # Full utterance text
'start_time': float, # Start time of the utterance (in seconds)
'end_time': float, # End time of the utterance (in seconds)
'words': [ # Word-level information
{
'word': str, # The word itself
'start_time': float, # Word-level start time
'end_time': float, # Word-level end time
}
]
}
],
'facial_expression': [ # Facial expression features
{
'utt_id': int, # ID of the utterance this expression is aligned to
'frame': int, # Frame identifier
'features': [float], # Facial feature vector (153-dimensional)
}
],
'body_language': [ # Body language features
{
'utt_id': int, # ID of the utterance this body language is aligned to
'frame': int, # Frame identifier
'features': [float], # Body movement feature vector (179-dimensional)
}
],
'speaker_bbox': [ # speaker bounding boxes
{
'frame_id': int, # Frame identifier
'bbox': [int], # [x_top, y_top, x_bottom, y_bottom]
}
],
'harmful_utterance_id': [int], # List of utterance IDs identified as harmful
}
Data Instances
See above
Data Fields
See above
Data Splits
Data splits can be accessed as:
from datasets import load_dataset
train_dataset = load_dataset("winston1214/VENUS-10K", split = "train")
test_dataset = load_dataset("winston1214/VENUS-10K", split = "test")
Curation Rationale
Full details are in the paper.
Source Data
We retrieve natural videos from YouTube and annotate the FLAME and SMPL-X parameters from EMOCAv2 and OSX.
Initial Data Collection
Full details are in the paper.
Annotations
Full details are in the paper.
Annotation Process
Full details are in the paper.
Who are the annotators?
We used an automatic annotation method, and the primary annotator was Youngmin Kim, the first author of the paper.
For any questions regarding the dataset, please contact e-mail
Considerations for Using the Data
This dataset (VENUS) consists of 3D annotations of human subjects and text extracted from conversations in the videos. Please note that the dialogues are sourced from online videos and may include informal or culturally nuanced expressions. Use of this dataset should be done with care, especially in applications involving human-facing interactions.
Licensing Information
The annotations we provide are licensed under CC-BY-4.0.