action-lab/gcaption
收藏资源简介:
GCaption是一个为多个RGB步态数据集提供的自然语言属性标注数据集,由GaitMax(CVPR26)引入。该数据集包含103,124个行走序列,覆盖四个子数据集:casiab、ccgrm、ccpg和sustech1k。每个序列通过7个属性进行描述(详见SCHEMA.md),并为每个属性提供预计算的OpenCLIP文本嵌入,以支持上下文感知的步态研究和条件去相关损失(CDLoss)。嵌入维度为(103124, 7, 768)的fp16格式。数据集仅包括文本标题和每属性的OpenCLIP嵌入,按序列ID键控,不包含源帧、掩码或姿态数据。标注基于Gemini-2.5-flash-lite和GPT-4o生成,嵌入使用OpenCLIP ViT-L-14模型计算。需要注意的是,属性(尤其是年龄)是模型推断的表观属性,未经地面真实验证,不应用于个体识别;源数据已去识别化,不包含个人身份信息。
Natural-language attribute annotations for multiple RGB gait datasets, introduced in GaitMax (CVPR26). Each walking sequence is described by 7 attributes (see SCHEMA.md) plus a precomputed OpenCLIP text embedding per attribute, to support context-aware gait research and the Conditional Decorrelation Loss (CDLoss). The dataset includes 103,124 sequences across four datasets: casiab, ccgrm, ccpg, and sustech1k, with embeddings of shape (103124, 7,768) in fp16 format. It provides text captions and per-attribute OpenCLIP embeddings keyed by sequence ID, but does not include source frames, masks, or poses. Captions are generated using Gemini-2.5-flash-lite and GPT-4o, and embeddings are computed with OpenCLIP ViT-L-14. Note that attributes (especially age) are model-inferred apparent attributes, not verified ground truth, and should not be used for individual identification; source data is de-identified and contains no PII.




