HACS（Human Action Clips and Segments）

HACS代表人类动作片段和片段，这是一个用于人类动作识别的视频数据集。

此数据集包含200个动作类别，与ActivityNet-v1.3数据集的分类相同。视频来自YouTube，大小为504K，每个视频不到4分钟，平均长度为2.6分钟。作者使用基于均匀随机性和图像分类器一致性/不一致性的方法对总共1.5M个视频片段进行采样，持续时间为2秒。0.6M的大小标记为阳性样品，0.9M的大小标记为阴性样品。。该数据集包括1.4M训练集，50k验证集和50k测试集，分别从492K，6k和6k视频中采样

HACS stands for Human Action Clips and Segments, a video dataset for human action recognition. It features 200 action categories consistent with those of the ActivityNet-v1.3 dataset. The videos are sourced from YouTube, comprising 504K total videos, each shorter than 4 minutes with an average duration of 2.6 minutes. The authors sampled a total of 1.5 million 2-second video clips using a method based on uniform randomness and the consistency/inconsistency of image classifiers. Among these, 0.6 million are labeled as positive samples, while 0.9 million are labeled as negative samples. This dataset includes a 1.4M training set, a 50k validation set, and a 50k test set, which are sampled from 492K, 6k, and 6k videos respectively.

搜集汇总

背景与挑战

背景概述

HACS是一个人类动作识别视频数据集，包含200个动作类别和504K个YouTube视频片段，平均时长2.6分钟。数据集分为1.4M训练集、50k验证集和50k测试集，适用于动作识别研究。