MPOSE2021 (MPOSE2021 Dataset for Short-time Human Action Recognition)

MPOSE2021 是作为 MPOSE 数据集 [1-3] 的演变而开发的。它是由 OpenPose [4] 和 Posenet [11] 在流行的 HAR 数据集上检测到的人体姿态数据制成的，即 Weizmann [5]、i3DPost [6]、IXMAS [7]、KTH [8]、UTKinetic-Action3D (RGB仅）[9] 和 UTD-MHAD（仅 RGB）[10]，以及原始视频数据集，即 ISLD 和 ISLD-Additional-Sequences [1]。由于这些数据集具有异构的动作标签，因此每个数据集标签都被重新映射到包含 20 个动作的共同且同质的列表。
该存储库允许用户直接访问 POSE 数据集（A 节）或以 Python 友好格式（B 节）为 MPOSE2021 生成 RGB 和 POSE 数据。生成的 RGB 和 POSE 序列的帧数在 20 到 30 之间。序列是通过切割所谓的“前体视频”（来自上述数据集的视频）获得的，具有不重叠的滑动窗口。 OpenPose/PoseNet 无法检测到任何主体的帧会被自动丢弃。结果样本当时包含一个受试者，执行单个动作的一小部分。
总体而言，MPOSE2021 包含 15429 个样本，分为 100 个受试者执行的 20 个动作。

MPOSE2021 was developed as an evolution of the MPOSE dataset [1–3]. It is constructed from human pose data detected by OpenPose [4] and Posenet [11] on popular HAR datasets, specifically Weizmann [5], i3DPost [6], IXMAS [7], KTH [8], UTKinetic-Action3D (RGB only) [9], and UTD-MHAD (RGB only) [10], along with the original video datasets ISLD and ISLD-Additional-Sequences [1].
Given that these datasets possess heterogeneous action labels, the labels from each individual dataset were remapped to a unified, homogeneous list encompassing 20 actions.
This repository allows users to either directly access the POSE datasets ("Section A") or generate RGB and POSE data for MPOSE2021 in a Python-friendly format ("Section B"). The frame counts of the generated RGB and POSE sequences fall between 20 and 30. These sequences are acquired by cropping the so-called "precursor videos" (videos sourced from the aforementioned datasets) via non-overlapping sliding windows. Frames in which no human subject was detected by OpenPose or PoseNet are automatically discarded. The final samples each contain a single subject performing a short segment of one individual action.
In total, MPOSE2021 comprises 15,429 samples, which are categorized into 20 actions executed by 100 distinct subjects.

搜集汇总

背景与挑战

背景概述

MPOSE2021是一个短时人体动作识别姿态数据集，包含15429个样本，涉及100个受试者执行的20个动作。该数据集基于多个流行的人体动作识别数据集和原始视频数据集构建，通过OpenPose和PoseNet检测人体姿态数据，并将异构的动作标签重新映射到共同列表。