Pixel Shift를 이용한 딥러닝 기반 영상 분할의 성능 개선 방법

Abstract

For image segmentation, both spatial information and contextual information are very important. CNN(Convolutional Neural Network) can make rich contextual information with deep layers. However,it is very difficult to preserve spatial information when we use CNN. We propose a novel method to preserve spatial information by using pixel shifted images. Using 4 pixel shifted images as inputs, the proposed method makes high-resolution feature maps. So, it has better performance for thin or small objects. It is very easy to attach to various networks because it does not need additional training. We show the performance improvement for semantic segmentation on Cityscapes dataset in the experimentation.