Kinetics-700

## 🎬 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.
- 🔬 **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.

🎬 Kinetics-700 数据集卡片

📦 🚨Kinetics-700 数据集解压须知🚨

若要充分使用Kinetics-700数据集，您需下载并解压全部22个压缩归档文件，此步骤为获取完整视频库的必要前提。若未完成全部归档的解压操作，将导致数据集不完整。

📝 数据集概述

Kinetics-700是面向人类动作识别（Human Action Recognition）的大规模YouTube视频URL数据集，是原始Kinetics数据集的扩展版本，大幅扩充了动作类别与视频片段数量。该数据集旨在为复杂多样的人类动作模型评估提供鲁棒基准。每个视频片段时长约10秒，仅展示单一人为动作。

本数据集共涵盖700个人类动作类别，囊括现实世界中丰富多样的活动类型。视频片段均取自各类YouTube视频，确保场景、拍摄主体与拍摄角度具有广泛分布性。

🏗️ 数据集结构

Kinetics-700数据集通常按训练集、验证集与测试集划分。数据集中的每条条目一般包含以下内容：
- 🎞️ **视频片段**：短视频片段（例如MP4格式）。
- 🏷️ **标签**：对应视频片段的动作类别（例如“拉小提琴”“骑自行车”）。
- 📜 **元数据（Metadata）**：相关附属信息，如原始YouTube视频ID以及片段的起始/结束时间戳。

由于授权与文件大小限制，视频文件本身通常不会直接分发。取而代之的是，数据集一般会提供YouTube视频ID列表与时间戳标注，用户可据此下载或提取特定视频片段。

📊 数据集统计信息
- 🔢 **类别数量**：700个独特人类动作类别。
- 💾 **数据集总大小**：871 GB
- 🎬 **总视频片段数**：635,000个视频片段。
- 🏋️ **训练集**：536,499个片段（约757 GB）
- ✅ **验证集**：33,966个片段（约35 GB）
- 🧪 **测试集**：64,535个片段（约79 GB；测试集标签未公开发布，以保证评测公平性）
- ⏱️ **平均片段时长**：每个片段约10秒。

✍️ 数据收集与标注流程

Kinetics-700的视频片段均从YouTube视频中提取。筛选流程由人类标注员完成，他们会从预定义的动作类别列表中匹配对应片段。针对每个动作类别，会收集多个视频示例以覆盖多样变化。标注信息会明确标注动作在源YouTube视频中的精确起始与结束时间。

💡 应用场景

Kinetics-700主要用于以下场景：
- 🏃‍♀️ **动作识别（Action Recognition）**：训练并评估用于视频中人类动作分类的深度学习模型（Deep Learning Model）。
- 🧠 **视频理解（Video Understanding）**：推动视频动态内容理解相关研究的发展。
- 🔄 **迁移学习（Transfer Learning）**：对大规模视频模型进行预训练，后续可针对其他视频相关任务进行微调。
- 🏆 **基准测试（Benchmarking）**：作为标准基准数据集，用于对比不同视频分析算法的性能表现。

🚧 局限性与偏差
- 🌐 **YouTube来源偏差**：由于数据源自YouTube，可能会反映用户生成内容中存在的固有偏差，例如主流文化活动或热门主题。
- 🗺️ **地域与人口统计偏差**：本数据集可能无法完全代表全球范围内人类动作的多样性，或存在向YouTube平台上占主导的地区与人口群体倾斜的情况。
- ❓ **动作定义模糊性**：尽管已定义动作类别，但部分动作可能存在细微变体或重叠，可能导致标注存在潜在歧义。
- ⚖️ **版权与授权问题**：使用YouTube内容需遵守YouTube服务条款及相关版权法规。
- 🔬 **仅限科研使用**：Kinetics数据集由其开发者提供，仅用于学术研究与非商业用途。用户需确保其应用符合上述条款，并注明原始来源。

ℹ️ 补充说明

使用Kinetics-700的用户需注意：由于数据集体积庞大，下载视频内容需要稳定的高速网络连接与充足的存储空间。数据集开发者通常会提供相关工具与脚本以简化该流程。使用视频内容时，您必须遵守YouTube服务条款及相关版权法规。

搜集汇总

背景与挑战

背景概述

Kinetics-700是一个包含700个动作类别、635,000个视频片段的大规模人类动作识别数据集，主要用于视频分析和深度学习模型训练。数据集来源于YouTube，需注意其潜在偏见和使用时的法律合规性。