Semi-supervised Synthetic-to-Real Domain Adaptation for Fine-grained Naval Ship Image Classification

EasyChair Preprint 2661

Abstract

In this paper, we propose a deep learning-based approach for fine-grained naval ship image classification. To this end, we tackle following two major challenges. First, to overcome the lack of the amount of training images in the target (i.e., real) domain, we generate a large number of synthetic naval ship images by using a simulation program which is specifically designed for our task. Second, to relieve performance degradation caused by the disparity between the synthetic and the real domains, we propose a novel regularization loss, named cross-domain triplet loss. Experimental results show that both the synthetic images and the proposed cross-domain triplet loss are essential to achieve the state-of-the-art performance for fine-grained naval ship image classification.

Keyphrases: Domain Adaptation, deep learning, fine-grained classification

@booklet{EasyChair:2661,
  author    = {Yoonhyung Kim and Hyeonjin Jang and Sangtae Park and Jiwon Lee and Changick Kim},
  title     = {Semi-supervised Synthetic-to-Real Domain Adaptation for Fine-grained Naval Ship Image Classification},
  howpublished = {EasyChair Preprint 2661},
  year      = {EasyChair, 2020}}