A self-supervised pure-vision three-dimensional target object automatic labeling method and device

Abstract

The application discloses a kind of self-supervised pure vision three-dimensional target object automatic labeling method and device, belong to computer vision field.The method mainly includes: single frame labeling and multi-frame fusion labeling.Single frame labeling first obtains pose initial estimation value by angle initial estimation model and two-dimensional bounding box comparison iterative algorithm.Picture after semantic segmentation and pose initial estimation value are input into encoder-decoder framework to obtain the three-dimensional expression of object, and the three-dimensional pose value of target object is obtained by angle sampling and pose iteration.Multi-frame fusion labeling combines the motion state of target object in multiple frames by judging the pose value of camera itself.The labeling result of entire video stream is inferred by close-range labeling result and the pose value of static object, and the labeling result of entire video stream is inferred by kinematics combined with dynamic object.Not only good labeling effect is achieved in self-supervised scheme, but also various camera parameters can be self-adapted, and manpower and material resources are saved in large-scale application.

Images