Learning methods and devices to allow CNNs learned in the virtual world used in the real world by converting run-time inputs using photo style conversion, and test methods and devices using them.

摘要

To allow a CNN having trained in a virtual world to be used in a real world.SOLUTION: A leaning method includes steps of: a learning device acquiring first learning images of virtual driving of a virtual vehicle; and the learning device performing a first learning process of instructing a main CNN to generate first estimated autonomous driving source information by referencing the first learning images, instructing the main CNN to generate first main losses by referencing first ground-truth autonomous driving source information corresponding to a first estimation and perform backpropagation using the first main losses, to thereby learn parameters of the main CNN, and a second learning process of instructing a supporting CNN to generate second learning images by referencing images of a first base corresponding to the first learning images and images of a second base of real driving of a real vehicle, instructing the supporting CNN to generate second estimated autonomous driving source information, instructing the supporting CNN to generate second main losses by referencing second ground-truth autonomous driving source information corresponding to a second estimation, and instructing the supporting CNN to perform backpropagation using the second main losses, to thereby learn parameters of the main CNN.SELECTED DRAWING: Figure 2