A Study of Zero-Cost Proxies for Remote Sensing Image Segmentation

Chen Wei, Tai Kai Guo, Ping Yi Tang, Yao Jun Ge, Jimin Liang

Abstract: Zero-cost proxies neural architecture search (NAS) can efficiently evaluate the performance of neural architectures and significantly reduce the search cost, but existing zero-cost proxies NAS are mainly focus on image classification. This paper investigates whether zero-cost proxies can accurately rank neural architectures used for remote sensing image segmentation. Firstly, we design a new search space for remote sensing image segmentation, denoted as SEG101, which considers enhancing the feature maps' contextual information and improving the fusion of feature maps. Secondly, a predictor-based NAS algorithm is adopted to explore SEG101 and collect neural architectures from it. Finally, zero-cost proxies are analysised by using the collected neural architectures. The preliminary experimental results illustrate that SEG101 is a promising search space and also show that zero-cost proxies can be used by predictor-based NAS for remote sensing image segmentation.

Keywords: Zero-Cost Proxies, Search Space, Semantic Segmentation, Remote Sensing

One-sentence Summary: In this paper, we designed a search space for remote sensing segmentation, verified the effectiveness of the search space, and investigated the characteristic of zero-cost proxies by using the architectures collected from the proposed search space.

Reproducibility Checklist: Yes

Broader Impact Statement: Yes

Paper Availability And License: Yes

Code Of Conduct: Yes

Reviewers:

<Chen Wei>, <[e-mail]>
<JiMin Liang>, <[e-mail]>