Transfer Learning using Transformation: Is Large Unlabeled Data Helpful at Segmentation?

Abstract

We propose a simple method of transfer learning for image segmentation. Creating labeled data for deep neural network training in image segmentation is particularly expensive than other tasks. Hence, practically, the labeled data is much less than the unlabeled data. So, we introduce a method that is helpful for segmentation by using unlabeled data. Our key is the RGB-to-HSV transformation and we use it in two ways. The first way is to pre-train a network to work as an RGB-to-HSV transformer which can extract useful features, and transfer the pre-trained weights to another network for segmentation, which is one of the most common transfer learning method. The second way is to provide additional information to the segmented network by providing HSV, the output of the pre-trained network, as additional input. We performed several experiments about our proposal using Cityscapes dataset.