CONVERSE

CONVERSE是一个人类互动识别数据集，旨在探索通过使用基于姿势和外观的特征对一对个体之间自然执行的对话场景进行分类。对行动识别社区的公开数据集的当前状态进行的深入调查强调了人类行动中的几个关键层，从单人行动到多人互动。当前基于深度的数据集描述了动作或交互，其中类通常由一系列关键姿势表示，这些姿势很容易在基于深度的表示 (即踢和拳) 中识别。CONVERSE背后的动机是提出使用基于姿势的信息对参与者之间的微妙和复杂行为进行分类的问题，这些类别不容易由其包含的姿势定义。通过使用商业深度传感器，记录了一对个体在7种不同的对话场景中进行自然对话，以提取的人体骨骼模型的形式提供基于姿势的交互表示。基线分类结果显示在相关出版物中，以允许与基于姿势的交互识别的未来研究进行交叉比较。CONVERSE is a human interaction recognition dataset aimed at exploring the classification of naturally occurring conversational scenarios between pairs of individuals using pose-based and appearance-based features. An in-depth survey of the current state of publicly available datasets in the action recognition community highlights several key layers of human actions, ranging from single-person actions to multi-person interactions. Current depth-based datasets describe actions or interactions, where categories are typically represented by a set of key poses that are readily identifiable in depth-based representations, such as kicking and punching. The core motivation behind CONVERSE is to address the problem of classifying subtle and complex interpersonal behaviors using pose-based information, categories that cannot be easily defined by their constituent poses. Using commercial depth sensors, we recorded natural conversations between pairs of individuals across 7 distinct conversational scenarios, with pose-based interactive representations provided in the form of extracted human skeletal models. Baseline classification results are presented in the associated publication to enable cross-comparison with future research on pose-based interaction recognition.

搜集汇总

背景与挑战

背景概述

CONVERSE是一个人类互动识别数据集，旨在利用姿势和外观特征对一对个体之间的自然对话场景进行分类。该数据集通过深度传感器记录了7种不同对话场景，并以骨骼模型形式提供基于姿势的交互表示，包含基线分类结果以支持未来研究比较。