HMDB51 Dataset

Dataset Summary

HMDB51 is a large video database for human motion recognition. It contains 6,766 video clips from 51 action categories, each containing a minimum of 101 clips. The actions can be grouped into five types:

1. General facial actions: smile, laugh, chew, talk
2. Facial actions with object manipulation: smoke, eat, drink
3. General body movements: cartwheel, clap hands, climb, climb stairs, dive, fall on the floor, backhand flip, handstand, jump, pull up, push up, run, sit down, sit up, somersault, stand up, turn, walk, wave
4. Body movements with object interaction: brush hair, catch, draw sword, dribble, golf, hit something, kick ball, pick, pour, push something, ride bike, ride horse, shoot ball, shoot bow, shoot gun, swing baseball bat, sword exercise, throw
5. Body movements for human interaction: fencing, hug, kick someone, kiss, punch, shake hands, sword fight

Supported Tasks

• Action Recognition: The dataset is designed for video-based human action recognition.

Dataset Structure

Data Fields

• video_id: Unique identifier for each video
• filename: MP4 video filename
• label: Action category (51 classes)
• original_filename: Original AVI filename
• duration: Video duration in seconds
• fps: Frames per second
• frame_count: Total number of frames
• width: Video width in pixels
• height: Video height in pixels
• resolution: Video resolution (width x height)
• file_size_original: Original AVI file size in bytes
• file_size_mp4: Converted MP4 file size in bytes
• split: Dataset split (train/validation/test)

Data Splits

• Train: 70% of videos per class
• Validation: 10% of videos per class
• Test: 20% of videos per class

Usage

import pandas as pd
from pathlib import Path

# Load metadata
train_metadata = pd.read_csv("train/metadata.csv")
validation_metadata = pd.read_csv("validation/metadata.csv")
test_metadata = pd.read_csv("test/metadata.csv")

# Get video paths and labels
train_videos = train_metadata['filename'].tolist()
train_labels = train_metadata['label'].tolist()

# Example: Load a specific video
video_path = Path("train") / train_metadata.iloc[0]['filename']
label = train_metadata.iloc[0]['label']

Source Data

Initial Data Collection and Normalization

The dataset was collected from various sources including movies, public databases, and YouTube videos. Videos were manually annotated with action labels.

Annotations

Each video is annotated with one of 51 action categories. The annotations were performed manually by human annotators.

Personal and Sensitive Information

This dataset contains videos of human actions. Users should be aware of privacy considerations when using this dataset.

Citation Information

@article{Kuehne2011HMDBAL,
  title={HMDB: A Large Video Database for Human Motion Recognition},
  author={Hilde Kuehne and Hueihan Jhuang and Estíbaliz Garrote and Tomaso Poggio and Thomas Serre},
  journal={2011 International Conference on Computer Vision},
  year={2011},
  pages={2556-2563}
}

Dataset Curators

The dataset was curated by the Serre Lab at Brown University.

Licensing Information

This dataset is provided for research purposes only. Please refer to the original dataset repository for licensing information.