Yixin Chen, Yaowei Zhang, Huangyue Yu, Junchao He, Yan Wang, Jiangyong Huang, Hongyu Shen, Junfeng Ni, Shaofei Wang, Baoxiong Jia, Song-Chun Zhu, Siyuan Huang

Researchers demonstrate that unlabeled internet videos can be automatically processed into training data for 3D scene understanding tasks, reducing reliance on expensive human annotation. The work uses carefully designed data engines to extract signal from web video, validating the approach across three task categories ranging from object detection to spatial reasoning. Models trained on this generated data show strong zero-shot performance and improve further with finetuning, suggesting a viable path to scaling 3D perception systems without proportional increases in annotation costs.