cycling
收藏资源简介:
Cycling World Model Dataset(自行车世界模型数据集)是一个专为视频世界模型研究设计的轻量级数据集。它包含一系列短的第一人称自行车骑行视频片段,这些片段由安装在自行车车把上的摄像头拍摄,真实地呈现了骑手在真实世界环境(特别是郊区)中的行进视角。数据集的主要目的是用于训练和评估能够根据过去观察预测未来视频帧的视频世界模型。针对现有视频数据集往往体积庞大、难以在消费级硬件上运行的问题,本数据集提供了一个更易于处理的研究替代方案。数据来源于在同一郊区进行的多次自行车骑行的较长录像,经过提取后形成独立的8秒片段。虽然每次骑行的具体路线有所变化,但许多地点会重复出现,这使得模型能够从略微不同的运动轨迹和不同的环境条件下学习同一场景的表示,有助于研究模型的泛化与条件预测能力。数据集总计包含2小时26分钟的视频内容,每个视频片段时长固定为8秒,由64帧组成,以8帧每秒(FPS)的帧率录制,视频分辨率为192像素×108像素。该数据集适用于计算机视觉领域的视频预测、世界模型构建、第一人称(自我中心)视觉理解等任务。
Cycling World Model Dataset is a lightweight dataset specifically designed for video world model research. It comprises a series of short first-person bicycle riding video clips captured by cameras mounted on bicycle handlebars, which authentically present the rider’s traveling perspective in real-world environments, especially suburban areas. The core objective of this dataset is to train and evaluate video world models capable of predicting future video frames based on past observations. To address the issue that existing video datasets are often excessively large and difficult to run on consumer-grade hardware, this dataset offers a more tractable research alternative. The data is sourced from longer recordings of multiple bicycle rides conducted in the same suburban region, and has been extracted into independent 8-second clips. While the specific routes of each ride differ, many locations are revisited, enabling models to learn representations of the same scene from slightly varied movement trajectories and different environmental conditions, which supports research on model generalization and conditional prediction capabilities. The dataset contains a total of 2 hours and 26 minutes of video content. Each video clip has a fixed duration of 8 seconds, consisting of 64 frames, recorded at a frame rate of 8 frames per second (FPS), with a resolution of 192 × 108 pixels. This dataset is applicable to tasks including video prediction, world model construction, and first-person (egocentric) visual understanding in the field of computer vision.




