Domain Adaptation For Semantic Segmentation

In recent years,deep learning has brought many new breakthroughs to the field of computer vision,especially in the field of image semantic segmentation,the record of segmentation accuracy is constantly updated.However,the model training depends on the dataset with pixel level annotation,and the huge cost of labeling hampers the expansion of application scenarios.This paper focuses on the domain adaptation algorithms for image semantic segmentation.The main work is as follows:(1)We propose to use conditional random field the domain adaptation for image semantic segmentation.The convolution kernel of traditional convolution neural network usually has a large receptive field,so it will produce rough output when the feature map is obtained for pixel level classification.We propose the use of conditional random fields for refining the output of the segmented network,which may further help to improve the capability of the domain discriminator.The experimental results show that the proposed conditional random field method can improve the edge sawtooth effect of the overall migration effect,and improve the final accuracy significantly.(2)For domain adaptation of semantic segmentation,we propose to employ atrous convolution for the design of more powerful discriminator.We introduce a dilated convolution structure into the structure of the domain discriminator,which can effectively expand the receptive field of the convolution kernel without incurring additional training parameters.Experimental results show that the proposed dilated convolution domain discriminator method has a significant performance improvement on the domain adaptation of image semantic segmentation.(3)We propose a novel design of domain discriminator network based on spatial pyramid pooling and multi-scale fusion.We introduce the spatial pyramid pool structure to the conventional full convolution domain discriminator for removing the limitation on the fixed size of the network,and getting rich feature information at different depths of each layer.Experimental results show that the proposed domain discriminator has a significant performance improvement over the traditional one.