blanchon/opencs2_dataset

OpenCS2 - POV Renders 是一个基于《反恐精英2》游戏的第一人称视角（POV）渲染数据集，专为视频分类、强化学习和其他相关任务设计。数据集包含每个玩家每回合的同步视频（1280x720分辨率、32帧/秒、近无损H.264编码）、音频（基于玩家位置和方向的立体声混合）、输入数据（如按键、鼠标移动、视角角度）和世界状态（如玩家位置、速度、健康状态）。数据集以松散文件布局组织，每个POV回合有独立的目录，包含视频、预览视频和tick数据。此外，提供了多个Parquet表用于索引和事件过滤，如回合索引、比赛元数据、击杀事件、对决事件等。数据集规模包括165,270个POV回合（约2974.2小时POV视频）、16,527个回合、794场比赛和111,715次击杀。用户可以通过配置过滤数据，并流式加载媒体文件，支持大规模训练和事件分析。OpenCS2 - POV Renders is a first-person perspective (POV) rendering dataset based on the Counter-Strike 2 game, designed for video classification, reinforcement learning, and other related tasks. The dataset includes synchronized video (1280x720 resolution, 32 fps, near-lossless H.264 encoding), audio (stereo mixed from player position and orientation), input data (e.g., keys, mouse delta, view angles), and world state (e.g., player position, velocity, health) for each player per round. It is organized in a loose-file layout, with each POV round having its own directory containing video, preview video, and tick data. Additionally, multiple Parquet tables are provided for indexing and event filtering, such as round index, match metadata, kill events, duel events, and more. The dataset scale includes 165,270 POV rounds (~2974.2 hours of POV video), 16,527 rounds, 794 matches, and 111,715 kills. Users can filter data via configs and stream media files, supporting large-scale training and event analysis.