具方向性之無錨框物體檢測方法

Anchor-free Oriented Object Detection

李浩(Hao Li)

指導教授 ： 蘇志文

中原大學/電機資訊學院/資訊工程研究所/碩士(2022年)

https://doi.org/10.6840/cycu202200131

摘要

近年來，旋轉物體檢測因遙測空拍影像檢測、零售標籤辨識等方面的重要性，於電腦視覺領域受到越來越多的重視。然而，目前傳統的水平邊界框不易於檢測如航拍影像中排列密集且有方向性的道路車輛、抑或是長寬比例懸殊的零售標籤、橋梁等，因此需要有別於主流的物體檢測器，去預測出能順應物體方向、貼合物體範圍的可旋轉邊界框。
有鑑於此，本論文捨棄基於迴歸的角度預測方式，改以非主流的角度分類方式來進行檢測，避免因角度迴歸所帶來的邊界不連續等問題。然而，角度分類因類別數量導致預測層的參數量過多，為此我們以無錨框單階段物體檢測方法YOLOX為原型，修改添加其角度預測等機制降低上述問題的影響。此外，傳統的網路模型在預測時使用的非極大值抑制(Non-Maximum Suppression, NMS)後處理方法，無法適應旋轉邊界框，為了更準確的預測旋轉物體，我們也加入適合旋轉邊界框的後處理方法，藉以提升網路模型的性能。
為了驗證本論文方法的有效性，我們採用大型遙測航拍公共資料集DOTA來進行訓練與驗證等工作，透過大量實驗與分析證明了我們所提的方法，具有良好出色的表現。

關鍵字

空拍影像處理 ； 無錨框物體檢測 ； 旋轉物體檢測

並列摘要

In recent years, oriented object detection has received increasing attention in the field of computer vision due to the object detection task on aerial images and scene text detection. However, the traditional bounding box cannot really fit the dense and oriented on-road vehicles and objects with high aspect ratios, such as scene text or bridges. Therefore, an oriented object detector is needed to generate rotatable bounding boxes that can really fit the specific objects.
In view of the above, the prediction of object angles in this work is not achieved by regression, but by classification, to avoid the discontinuity of angle values that can be caused by regression. However, the number of parameters in the prediction layer is quite large. Therefore, an anchor-free one-stage object detector, YOLOX, is adopted for the baseline. We modified the branches and loss functions of YOLOX to compensate the impact of angle classification. In addition, we add a non-maximum suppression post-processing to make oriented object detection more accurately.
To verify the effectiveness of the proposed method, a large-scale public aerial image dataset, DOTA, is adopted for training and validation. Numerous experiments and analyses show the outstanding performance of our proposed method.

並列關鍵字

Aerial image processing ； Anchor Free Object detection ； Oriented Object detection

參考文獻

Zheng, G., Liu, S., Wang, F., Li, Z., Sun, J. (2021). YOLOX: Exceeding YOLO Series in 2021. Cornell University Library, arXiv.org.

Yang, X., Yan, J. (2020). Arbitrary-Oriented Object Detection with Circular Smooth Label. ECCV.

Redmon, J., Divvala, S.K., Girshick, R.B., Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788.

Redmon, J., Farhadi, A. (2017). YOLO9000: Better, Faster, Stronger. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6517-6525.

Redmon, J., Farhadi, A. (2018). YOLOv3: An Incremental Improvement. ArXiv, abs/1804.02767.