University of Leicester
Browse

Rotated Object Detector with Self- Attention and Improved IoU Loss

Download (2.66 MB)
conference contribution
posted on 2023-03-07, 15:46 authored by H Qian, N Li, N Yuan, Z Zhang, Huiyu Zhou

Nowadays in UAV and remote sensing image processing area, rotated object detection has been attached more and more attentions. However, there are few studies of this based on Transformer, which is much stronger than CNN for feature extraction. Moreover, the well-developed IoU based loss function in horizontal object detection area is not well fits with rotated objects. In this paper, we argue that Transformer of pyramid structure called Swin-Transformer is an effective alternative of CNN. Specifically, the finetuned Swin-Transformer with Relatively Position Encoding (RPE) performed much better than other backbones generally used. Moreover, the new kind of IoU loss called Gaussian Estimate Loss (GE Loss) that use gaussian kernel to model object is applied in our model. It can increase the precious of the model. This loss is in contract to other Gaussian modeling loss function for adding direction vector that can solve the difficulty of testing objects close to square. Experiments on DOTA dataset achieved 84.99% map, which indicates that and our experiment shows that these improvements of our model are effective.

History

Author affiliation

School of Engineering, University of Leicester

Source

International Symposium on Design Studies and Intelligence Engineering

Version

  • VoR (Version of Record)

Published in

Design Studies and Intelligence Engineering

Volume

365

Pagination

94-103

Publisher

IOS press

isbn

978-1-64368-373-7

Copyright date

2023

Available date

2023-03-07

Editors

Lakhmi C. Jain, Valentina Emilia Balas, Qun Wu, Fuqian Shi

Book series

Frontiers in Artificial Intelligence and Applications

Language

en

Usage metrics

    University of Leicester Publications

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC