Learn CNN-based object detectors applicable to user requirements, such as key performance indicators, using target object integration networks and target object prediction networks for use in multiple cameras or surround view monitoring. Method and learning device, and testing method and testing device using the same

摘要

PROBLEM TO BE SOLVED: To provide a method for learning parameters of a CNN-based object detector suitable for a user requirement such as a key performance evaluation index by using a target object integration network and an object area prediction network. SOLUTION: The CNN can be redesigned as the resolution or focal length related to the important performance evaluation index changes and the scale of the object changes. In the method, the learning apparatus 100 searches the k-th predicted target area with the target object prediction network, and generates the k-th to k-th object proposals corresponding to the objects on the k-th to k-th processed images with the RPN. And integrating the object proposal with the target object integration network and integrating the k_1 to k_n object detection information output from the FC layer. [Selection diagram] Figure 2