Fourier-FPN: Fourier Improves Multi-scale Feature Learning for Oriented Tiny Object Detection

Precise detection of tiny objects in oriented object detection remains a significant challenge due to their limited visual information and frequent occurrence within scenes. Concretely, extremely small in-stances occupy a large proportion in oriented object detection, and the local receptive field of CNN-based detectors leads to the extraction of limited CNN features, which are insufficient for the detection of tiny oriented objects. To alleviate the aforementioned issue, we propose a Fourier-Aware Feature Pyramid Network (Fourier-FPN), which is tailored for the detection of tiny oriented objects and is based on channel-dimension Fourier Transform. For one thing, we introduce the Context-aware Fourier Feature Learning Module (CFFLM), which ensures the availability of sufficient and context-aware Fourier features for tiny objects by aggregating limited CNN features from the backbone using channel-dimension Fourier transform. For another, we propose a Multi-scale Cross-type Feature Fusion Module (MCFFM) that fuses context-aware Fourier features into insufficient CNN features, thereby obtaining sufficient features capable of detecting small oriented objects. Instead of redesigning the feature extraction network, our approach directly learns sufficient features from insufficient tiny oriented object features, utilizing the global receptive field of the Fourier Transform. Experimental results demonstrate the effectiveness of our Fourier-FPN on standard oriented object detection datasets, particularly with small object in-stances such as DOTA-v1.5 and HRSC2016. Our code will be available at https://github.com/tangyongst/Fourier-FPN .