METHOD FOR DESIGNING DEFORMABLE CONVOLUTIONAL NEURAL NETWORKS AND APPARATUS USING THEREOF

摘要

A method for a computing device operated by at least one processor to train a convolutional neural network, the method comprising: determining offset information including a sampling direction and location of an input feature map; proportional to a maximum value of an offset included in the offset information generating a loss function using the size of the receiving region as a variable, sampling data of the input feature map based on the offset information to generate a new input feature map, and using the new input feature map and a convolution filter performing a convolution operation to generate an output feature map, and optimizing the convolutional neural network using the loss function.