Okutama-Action

用于鸟瞰并发人体动作检测的新视频数据集。它由 43 分钟长的全注释序列和 12 个动作类组成。 Okutama-Action 具有当前数据集中缺少的许多挑战，包括动作的动态转换、比例和纵横比的显着变化、突然的摄像机移动以及多标签演员。

A novel video dataset for bird's-eye view concurrent human action detection. It consists of 43 minutes of fully annotated sequences and 12 action classes. Okutama-Action features numerous challenges absent from current datasets, including dynamic transitions of actions, significant variations in scale and aspect ratio, abrupt camera movements, and multi-label actors.

搜集汇总

背景与挑战

背景概述

Okutama-Action是一个用于鸟瞰视角下并发人体动作检测的视频数据集，包含43分钟的全注释序列和12个动作类。该数据集具有动作动态转换、比例和纵横比显著变化、突然摄像机移动以及多标签演员等挑战性特点，适用于计算机视觉中的动作检测和目标跟踪任务。