IFDet: Interactive and Flexible Single-Stage Detector

Refined Single-Stage Detector with Feature Refinement for Rotating Objects (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>DET), renowned for its efficiency and precision, has emerged as a high-performance detector. Nevertheless, its performance is hindered by limitations in feature extraction from the backbone network and challenges posed by boundary issues. To overcome these obstacles, we introduce two key innovations:Attention-ResNet Backbone and InteractiveHead. We present a novel Attention-ResNet backbone that seamlessly integrates attention mechanisms into the ResNet architecture. This fusion not only amplifies the network’s feature extraction capabilities but also maintains computational efficiency, ensuring limited iterations.To tackle boundary problems, we devise the InteractiveHead, which employs Kalman filtering within the IoU loss function. This approach offers enhanced robustness and adaptability in handling object boundaries, leading to improved localization accuracy.Building upon these advancements, we propose a novel Interactive and Flexible Single-Stage Detector. Our detector, renowned for its stability and efficiency, achieves a notable 55.7% mAP on the DOTA1.0 dataset without relying on pretrained models. This represents a substantial 5.0% improvement over the R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>DET baseline, all while maintaining equivalent computational complexity. These results underscore the effectiveness of our proposed approach in elevating the performance of single-stage detectors for rotating objects.