A learning method and a learning device for improving segmentation performance used for detecting a road user event using a double embedding structure in a multi-camera system, and a testing method and a testing device using the same. {LEARNING METHOD AND LEARNING DEVICE FOR IMPROVING SEGMENTATION PERFORMANCE TO BE USED FOR DETECTING ROAD USER EVENTS USING DOUBLE EMBEDDING CONFIGURATION IN MULTI-CAMERA SYSTEM AND TESTING METHOD AND TESTING DEVICE USING THE SAME}

摘要

PROBLEM TO BE SOLVED: To provide a learning method for improving the performance of segmentation used for road user event detection in a multi-camera system. A learning method applies a similarity convolution operation to a feature output from a neural network to generate a similarity embedding feature, and the similarity between two points sampled from the similarity embedding feature and the corresponding value. Generate the similarity loss with reference to the GT label image, apply the distance convolution operation to the similarity embedding feature, generate the distance embedding feature, and increase the inter-class difference between the average values of the instance classes. Then, a distance loss is generated in order to reduce the variance value within the class of the instance class, and at least one of the similarity loss and the distance loss is back-propagated. [Selection diagram] Figure 2