Rotation Equivariant Feature Image Pyramid Network for Object Detection in Optical Remote Sensing Imagery

Detection of objects is extremely important in various aerial vision-based applications. Over the last few years, methods based on convolution neural networks (CNNs) have made substantial progress. However, because large variety object scales, densities, and arbitrary orientations, current detectors struggle with extraction semantically strong features for small-scale by a predefined kernel. To address this problem, we propose rotation equivariant feature image pyramid network (REFIPN), an equivariance convolution. The proposed model adopts single-shot detector parallel lightweight module (LIPM) to extract representative generate regions interest optimization approach. extracts wide range scales orientations using novel filters. These are used vector fields determine weight angle highest-scoring orientation all spatial locations image. By approach, performance small-sized detection enhanced without sacrificing large-sized detection. validated two commonly benchmarks results show our can achieve state-of-the-art satisfactory efficiency.