CS2-10k
收藏资源简介:
CS2-10k是一个大规模、以自我为中心(第一人称视角)的《反恐精英2》游戏数据集,基于从HLTV网站获取的公开职业比赛演示构建。数据集包含超过60万个玩家回合视频,总计超过1万小时的第一人称游戏画面,视频以720p分辨率和48帧每秒渲染。每个视频都配有精确同步的逐帧注释数据,存储为.parquet文件。注释分为两个层次:剪辑级元数据(包括地图名称、回合编号、玩家队伍、总帧数、帧率、时长和视野)和逐帧注释(记录玩家的按键操作、相机视角增量、三维坐标以及偏航角和俯仰角)。该数据集适用于需要第一人称视角时序行为数据的研究任务,如世界模型构建、模仿学习、动作预测、游戏AI智能体训练和计算机视觉中的以自我为中心的视频理解,采用CC BY-NC 4.0许可证发布,仅供研究用途。
CS2-10k is a large-scale, egocentric (first-person perspective) dataset for the game Counter-Strike 2, constructed using public professional match demos obtained from the HLTV website. The dataset contains over 600,000 player round videos, totaling more than 10,000 hours of first-person in-game footage rendered at 720p resolution and 48 frames per second. Each video is paired with precisely synchronized frame-by-frame annotation data stored in .parquet files. The annotations are split into two tiers: clip-level metadata (including map name, round number, player team, total frame count, frame rate, duration, and field of view) and frame-by-frame annotations, which record the player's input actions, camera view delta, three-dimensional coordinates, as well as yaw and pitch angles. This dataset is applicable to research tasks requiring first-person temporal behavioral data, such as world model construction, imitation learning, action prediction, game AI Agent training, and egocentric video understanding in computer vision. It is released under the CC BY-NC 4.0 license for research use only.




