首页>
外国专利>
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
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
A learning method for improving segmentation performance to be used to detect road user events including pedestrian events and car events using a double embedding configuration in a multi-camera system is provided. The learning method includes the steps of: (i) having a similarity convolution layer generate a similarity embedding feature by applying a similarity convolution operation to a feature output from a neural network; (ii) causing the similarity loss layer to generate a similarity loss by referring to the similarity between two points sampled from the similarity embedding feature and a GT label image corresponding thereto; (iii) causing a distance convolution layer to generate a distance embedding feature by applying a distance convolution operation to the similarity embedding feature; (iv) causing the distance loss layer to generate a distance loss to increase the inter-class difference between the mean values of the instance classes and decrease the intra-class variance of the instance classes; and (v) backpropagating at least one of the similarity loss and the distance loss.
展开▼
