JaneDing2025/simact_video

SimAct Video数据集包含用于短人类动作片段的前后图像对。该数据集从多个视频动作识别数据集中转换而来，采用共享的JSONL + tar文件布局。每个JSONL行指向两张图像：precon表示起始视觉状态，postcon表示结束视觉状态。图像路径相对于存储库根目录，并与tar文件块中的路径匹配。本版本包含543,005个示例和1,086,010张图像，涵盖6个数据集，图像被打包成35个tar文件块。数据集包括ego4d_fho_lta、fineaction、holoassist、hd_epic、assembly101和epic_kitchens_100，分别用于训练、验证和测试分割。JSONL文件包含数据集名称、分割类型、ID、原始ID、图像路径、来源以及原始动作标签（如自然语言句子、动词和名词列表）等字段。SimAct Video contains before/after image pairs for short human action segments, converted from multiple video action-recognition datasets into a shared JSONL + tar layout. Each JSONL row points to two images: precon is the starting visual state and postcon is the ending visual state. Image paths are relative to the repository root and match paths stored inside the tar chunks. This release contains 543,005 examples and 1,086,010 images across 6 datasets. Images are packaged into 35 tar chunks. The datasets include ego4d_fho_lta, fineaction, holoassist, hd_epic, assembly101, and epic_kitchens_100, with splits for training, validation, and testing. The JSONL files include fields such as dataset name, split type, ID, original ID, image paths, source, and raw action labels (e.g., natural-language sentence, verb and noun lists).