Evaluating Domain Translation Approaches for Drone-Based Geo-Localization in Adverse Weather
Yiqing Li, Shuke He, Jin Chen · 2024
The UAVs in Multimedia (UAVM) 2024 competition aims to improve the performance of Drone-based Geo-localization task under extreme weather. On the task, we found the simple augmentation on training set can significant improve performance with zero-cost and won the 6th Place among 22 teams on this competition. Furthermore, based on our observations, we propose a domain translation framework to further enhance the performance of any model. We transfer images from a multi-weather domain to a normal domain using GAN and introduce FFM to further reduce computational costs. Our experiments revealed that while this strategy indeed improves image quality, it does not translate into improved recognition accuracy. Thus, while our strategy is supported by experimental evidence, practical methods for achieving performance improvement remain elusive. We hope this report provides valuable insights for developers and researchers in this field.