EOOD: End-to-end oriented object detection

Caiguang Zhang, Zilong Chen, Boli Xiong, Kefeng Ji, Gangyao Kuang

Abstract: Recently, significant advancements have been made in oriented object detectors based on convolutional networks. However, these models often rely on hand-designed post-processing techniques such as non-maximum suppression (NMS) to suppress redundant predictions, which impedes establishing an end-to-end detection system. In this paper, we explore how to build an end-to-end oriented detector? Firstly, our research demonstrates that prediction-oriented one-to-one label assignment (POLA) can significantly reduce the performance gap between using and not using NMS and is an essential component of end-to-end detection. Additionally, the proposed Negative Sample Reweighted Focal Loss (NFL) can widen the classification confidence gap between positive and negative samples, separate positive samples from noise, and guarantee high classification scores for the single positive sample. Finally, in order to address the lack of supervision caused by one-to-one label assignment, a joint training pipeline is designed that unites multiple auxiliary heads and takes advantage of one-to-many label assignment on supervision to improve feature representation and increase performance. During inference, Only the main head trained with one-to-one label assignment is involved in prediction. Extensive experiments on publicly available datasets DOTA and DIOR-R demonstrate that the proposed EOOD exhibits significant performance improvements over baseline models and has the potential to overcome the limitations of NMS. Code is available at https://github.com/zhangiguang/EOOD.

External IDs:dblp:journals/ijon/ZhangCXJK25