Aerial Image Segmentation via Noise Dispelling and Content Distilling

Yongqing Sun, Xiaomeng Wu, Yukihiro Bandoh, Masaki Kitahara; Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops, 2022, pp. 265-275

Aerial image segmentation is an essential problem for land management which can be used for change detection and policy planning. However, traditional semantic segmentation methods focus on single-perspective images in road scenes, while aerial images are top-down views and objects are of a small size. Existing aerial segmentation methods tend to modify the network architectures proposed for traditional semantic segmentation problems, yet to the best of our knowledge, none of them focus on the noisy information present in the aerial images. In this work, we conduct an investigation on the effectiveness of each channels of the aerial image on the segmentation performance. Then, we propose a disentangle learning method to investigate the differences and similarities between channels and images, so that potential noisy information can be removed for higher segmentation accuracy.