公开(公告)号：US11461958B2

公开(公告)日：2022-10-04

申请号：US17216719

申请日：2021-03-30

Applicant: UBTECH ROBOTICS CORP LTD

Inventor： Xi Luo ,  Mingguo Zhao ,  Youjun Xiong

IPC: G06T15/20 ,  G06F30/20 ,  G06V20/10 ,  G06N3/08

Abstract: A scene data obtaining method as well as a model training method and a computer readable storage medium using the same are provided. The method includes: building a virtual simulation scene corresponding to an actual scene, where the virtual simulation scene is three-dimensional; determining a view frustum corresponding to preset view angles in the virtual simulation scene; collecting one or more two-dimensional images in the virtual simulation scene and ground truth object data associated with the one or more two-dimensional images using the view frustum corresponding to the preset view angles; and using all the two-dimensional images and the ground truth object data associated with the one or more two-dimensional images as scene data corresponding to the actual scene. In this manner, the data collection does not require manual annotation, and the obtained data can be used for training deep learning-based perceptual models.

公开(公告)号：US20220139027A1

公开(公告)日：2022-05-05

申请号：US17216719

申请日：2021-03-30

Applicant: UBTECH ROBOTICS CORP LTD

Inventor： Xi Luo ,  Mingguo Zhao ,  Youjun Xiong

IPC: G06T15/20 ,  G06K9/00 ,  G06F30/20 ,  G06N3/08

Abstract: A scene data obtaining method as well as a model training method and a computer readable storage medium using the same are provided. The method includes: building a virtual simulation scene corresponding to an actual scene, where the virtual simulation scene is three-dimensional; determining a view frustum corresponding to preset view angles in the virtual simulation scene; collecting one or more two-dimensional images in the virtual simulation scene and ground truth object data associated with the one or more two-dimensional images using the view frustum corresponding to the preset view angles; and using all the two-dimensional images and the ground truth object data associated with the one or more two-dimensional images as scene data corresponding to the actual scene. In this manner, the data collection does not require manual annotation, and the obtained data can be used for training deep learning-based perceptual models.