Method for generating target sample based on CG

Abstract

The invention discloses a method for generating a target sample based on CG. The method includes following steps: establishing an initial target sample model database and an initial target environmentsample model database; randomly extracting at least an initial target sample model from the initial target sample model database, and randomly extracting at least an initial target environment samplemodel from the initial target environment sample model database; combining the extracted initial target sample model with the extracted initial target environment sample model; adjusting parameters of the initial target environment sample model, and adjusting parameters of the initial target sample model; setting a visual field range parameter, a resolution parameter and a distortion parameter ofan analog camera for the extracted initial target sample model and the initial target environment sample model; performing rendering to form an image, calculating a 2D bounding box, and calculating a3D bounding box; and forming and outputting a target sample.

Description

technical field [0001] The invention relates to a method for generating target object samples based on CG, in particular to a method for producing targets such as vehicles, people, buildings, roads and the like in ADAS active safety systems based on CG technology. Background technique [0002] In the field of machine learning both databases and learning algorithms have a crucial impact on the resulting model. In the framework of deep learning, the size and quality of the database have an approximately linear impact on the resulting model, that is, the resulting model of deep learning is proportional to the size and quality of the training database. The scale of data used in deep learning is in the millions or even tens of millions. The workload of manual labeling is extremely huge, and it is difficult to guarantee the quality of labeling. In terms of Scene Depth, Segmentation, Dense Optical Flow, and Environmental Structure, the workload of manual labeling is even greater,...

AI Technical Summary

Problems solved by technology

It takes a long time for a single image and the labeling results are inconsistent. For databases with millions or even tens of millions of levels, this kind of dense labeling is almost impossible to do manually.

Images