The classically oriented object detection method often suffers boundary discontinuity and square-like problems, hindering the model's ability to predict orientation accurately. Therefore, in this letter, we introd...
详细信息
The classically oriented object detection method often suffers boundary discontinuity and square-like problems, hindering the model's ability to predict orientation accurately. Therefore, in this letter, we introduce a novel angle representation scheme named coupled dual-frequency phase-shifting coder (CDFP) which draws inspiration from optical measurement technology to address the aforementioned issues. Specifically, in the angle encoding stage, we represent the ground-truth (GT) rotation angle as a combination of two five-step phase-shifting at different frequencies and use this representation to supervise the learning of the model's angle branch. In the angle decoding stage, alongside utilizing the corresponding dual-frequency phase-shifting decoding and unwrapping method, we impose additional constraints on the decoding angle range for predicted square-like objects. Extensive experiments on three challenging aerial image datasets using different detectors prove the effectiveness of our approach. Specifically, our RetinaNet-CDFP achieves an average improvement of 2.16% AP50 and 6.83% AP75 on DOTA, and when combined with RTMDet, our RTMDet-R-m-CDFP achieves state-of-the-art (SOTA) detection performance on DIOR-R and DOTA, with 70.11% and 78.77% AP50, respectively. Our codes will be released at https://***/liufeinuaa/***.
暂无评论