AVA-ActiveSpeaker

包含视频中时间标记的人脸轨迹，其中每个人脸实例都被标记为说话与否，以及语音是否可听。该数据集包含大约 365 万个人类标记帧或大约 38.5 小时的面部轨迹，以及相应的音频。

This dataset contains time-stamped facial trajectories from videos. Each facial instance is annotated as either speaking or non-speaking, and whether the corresponding speech is audible. The dataset includes approximately 3.65 million human-annotated frames, or equivalently roughly 38.5 hours of facial trajectories, along with the corresponding audio.

搜集汇总

背景与挑战

背景概述

AVA-ActiveSpeaker是一个用于主动说话人检测的视听数据集，包含约365万标记帧(38.5小时)的视频人脸轨迹及对应音频，每个人脸实例都标注了说话状态。该数据集由Google AI Research于2019年发布，支持音频-视觉联合分析研究。