ybashir/CS2-HUD-OCR-Crops

CS2 HUD OCR Crops数据集是一个用于图像到文本任务的数据集，专门针对《反恐精英2》游戏的平视显示器区域进行光学字符识别。数据集包含从三场由同一玩家（farouqqq）录制的比赛录像中提取的HUD区域裁剪图像，标注信息主要来源于游戏录像的demo文件状态解析（parse_ticks状态），部分信号使用TesserOCR进行标注。该数据集旨在训练一个专用CRNN模型，以替代现有的EasyOCR击杀信息读取器，提升处理速度（目标为低于30毫秒）。数据集涵盖多种信号类型，如生命值、金钱、弹药、计时器、击杀镜头、击杀信息和友好聊天，其中击杀信息和友好聊天在此版本中未标注。数据集按样本和回合进行划分，避免跨分割帧泄露，文件结构包括PNG格式的裁剪图像和Parquet格式的数据表。数据集采用CC-BY-4.0许可证，用于非商业研究目的。CS2 HUD OCR Crops is a dataset for image-to-text tasks, specifically designed for optical character recognition on heads-up display regions from Counter-Strike 2 game recordings. It consists of per-region HUD crops sliced from three match recordings by the same player (farouqqq), labelled where possible from the demo files parse_ticks state, with some signals annotated using TesserOCR. The dataset is built to train a specialist CRNN that replaces an existing EasyOCR killfeed reader, aiming for sub-30 ms performance. It includes multiple signal types such as health points, money, ammo, timer, killcam, killfeed, and friendly chat, with killfeed and friendly chat unlabeled in this version. The dataset is split by sample and round to prevent frame leakage across splits, and the file layout comprises PNG crop images and Parquet data tables. It is licensed under CC-BY-4.0 for non-commercial research purposes.