it-just-works/shot-boundary-detection

本数据集支持研究视频片段边界检测（SBD），提供超过340万个结构统一的61帧视频片段。每个片段以关键帧（第31帧）为中心，被标记为：C（剪辑）：直接的视频片段边界，T（转换）：渐变的转换（例如，淡出，溶解），E（空）：没有边界。数据来自AutoShot、ClipShots和爬取的Pexels视频，包括真实和合成的示例。每个片段包括元数据字段，如label，origin，video_id和video路径。数据集分为训练集（83%）和测试集（17%），确保源视频之间没有重叠。预处理包括严格的窗口化、去重和质量检查。虽然全面，但用户应意识到可能存在标注噪声、合成偏差和FFMPEG伪影。推荐使用精确度、召回率、F1分数和混淆矩阵进行评估。

This dataset supports research in shot boundary detection (SBD) by providing over 3.4 million uniformly structured 61-frame video clips. Each clip is centered on a key frame (the 31st), labeled as C (Cut) for a direct shot boundary, T (Transition) for a gradual transition, or E (Empty) for no boundary present. Data is sourced from AutoShot, ClipShots, and crawled Pexels videos, with both real and synthetically generated examples. Each clip includes metadata fields like label, origin, video_id, and video path. The dataset is split into training (83%) and testing (17%) sets, ensuring no overlap in source videos. Preprocessing involved strict windowing, deduplication, and quality checks. While comprehensive, users should be aware of possible annotation noise, synthetic biases, and FFMPEG artifacts. Evaluation is recommended using Precision, Recall, F1-score, and confusion matrices.