EgoNet

EgoNet是由字节跳动研究创建的一个大规模数据集，包含近500,000个视频片段，覆盖数百种场景，并富含人与物体的交互。该数据集基于Ego4D构建，通过截取1秒的短视频片段并进行10倍下采样，最终形成约150万个视频帧。EgoNet旨在通过丰富的视觉特征，辅助机器人学习视觉预训练，特别是在机器人操作任务中，通过自监督和监督学习相结合的方式，提升机器人在复杂环境中的操作能力。

EgoNet is a large-scale dataset created by ByteDance Research. It contains nearly 500,000 video clips covering hundreds of scenarios and rich in human-object interactions. Built upon Ego4D, this dataset is processed by extracting 1-second short video clips and performing 10-fold downsampling, ultimately yielding approximately 1.5 million video frames. EgoNet aims to facilitate visual pre-training for robots via its abundant visual features; specifically, by combining self-supervised and supervised learning approaches, it enhances robots' operational capabilities in complex environments for robotic manipulation tasks.