The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Error code: JobManagerCrashedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
video
video | label
string | youtube_id
string | start_time
int64 | end_time
int64 |
|---|---|---|---|---|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
|
null | null | null | null |
🎬 Dataset Card for Kinetics-700
📦 🚨IMPORTANT Dataset Decompression for Kinetics-700🚨
To fully utilize the Kinetics-700 dataset, you must download and decompress all 22 zipped archives. This process is essential to access the complete video collection. Failure to decompress all archives will result in an incomplete dataset.
📝 Dataset Description
The Kinetics-700 dataset is a large scale collection of YouTube video URLs for human action recognition. It is an extension of the original Kinetics dataset, significantly increasing the number of action classes and video clips. The dataset aims to provide a robust benchmark for evaluating models on complex and diverse human actions. Each video clip is approximately 10 seconds long and depicts a single human action.
This dataset is categorized into 700 human action classes, encompassing a broad array of real world activities. The clips are drawn from various YouTube videos, ensuring a wide distribution of settings, subjects, and camera angles.
🏗️ Dataset Structure
The Kinetics-700 dataset is typically structured with distinct splits for training, validation, and testing. Each entry in the dataset generally includes:
- 🎞️ Video Clip: A short segment of video (e.g., MP4 format).
- 🏷️ Label: The corresponding action class for the video clip (e.g., "playing violin," "riding a bicycle").
- 📜 Metadata: Associated information such as the original YouTube video ID and the start/end timestamps of the clip.
The video files themselves are not typically distributed directly due to licensing and size constraints. Instead, the dataset commonly provides a list of YouTube video IDs and timestamp annotations, from which users can download or extract the specific clips.
📊 Dataset Statistics
- 🔢 Number of Classes: 700 unique human action classes.
- 💾 Total Dataset Size: 871 GB
- 🎬 Total Video Clips: 635,000 video clips.
- 🏋️ Training Set: 536,499 clips (approximately 757 GB)
- ✅ Validation Set: 33,966 clips (approximately 35 GB)
- 🧪 Test Set: 64,535 clips (approximately 79 GB; labels not publicly released for fair comparison)
- ⏱️ Average Clip Duration: Approximately 10 seconds per clip.
✍️ Data Collection and Annotation
The video clips for Kinetics-700 are extracted from YouTube videos. The selection process involves human annotators who identify specific segments corresponding to a predefined list of action categories. For each action class, multiple video examples are collected to capture variability. The annotations specify the precise start and end times within the source YouTube video where the action occurs.
💡 Use Cases
Kinetics-700 is primarily used for:
- 🏃♀️ Action Recognition: Training and evaluating deep learning models to classify human actions in video.
- 🧠 Video Understanding: Advancing research in understanding dynamic content in video.
- 🔄 Transfer Learning: Pre-training large-scale video models that can then be fine-tuned for other video related tasks.
- 🏆 Benchmarking: Serving as a standard benchmark for comparing the performance of different video analysis algorithms.
🚧 Limitations and Biases
- 🌐 YouTube Source Bias: As the data is sourced from YouTube, it may reflect biases present in user generated content, such as dominant cultural activities or popular themes.
- 🗺️ Geographical and Demographic Skew: The dataset might not perfectly represent the diversity of human actions globally, potentially showing a skew towards regions or demographics prevalent on YouTube.
- ❓ Action Definition Ambiguity: While action classes are defined, some actions can have subtle variations or overlap, leading to potential ambiguities in labeling.
- ⚖️ Copyright and Licensing: The use of YouTube content necessitates adherence to YouTube's terms of service and any applicable copyright laws. The provided dataset typically includes only URLs and annotations, not the video content itself, shifting the responsibility of video download to the user.
- 🔬 Research Use Only: The Kinetics dataset is provided by its creators for academic research and non commercial use. Users should ensure their application aligns with these terms and acknowledge the original source.
ℹ️ Additional Information
Users intending to work with Kinetics-700 should be aware that downloading the video content requires a robust internet connection and significant storage due to its large size. Tools and scripts are often provided by the dataset creators to facilitate this process. You must comply with YouTube's Terms of Service and any relevant copyright laws when using the video content.
- Downloads last month
- 94