Reproduce “Sparse R-CNN OBB: Ship Target Detection in SAR Images Based on Oriented Sparse Proposals” on OpenResearch

About this paper

Kamirul Kamirul, Odysseas Pappas, Alin Achim

This research from the University of Bristol introduces Sparse R-CNN OBB, a deep learning framework that utilizes sparse learnable proposals for oriented ship detection in Synthetic Aperture Radar (SAR) images. The model simplifies the detection pipeline by eliminating dense anchors and achieves an overall Average Precision (AP50) of 91.82% on the RSDD-SAR dataset, notably reaching 96.26% AP50 in offshore environments.

Key insights

• Sparse learnable proposals can be effectively extended to handle oriented object detection in SAR images, providing an architectural simplification compared to dense anchor-based methods while maintaining high performance.
• The use of Rotated RoIAlign and modified 5-parameter regression layers are crucial for accurately capturing and refining arbitrary object orientations within a sparse proposal framework.
• A balance exists between the number of sparse proposals and computational cost; 300 proposals provided the best trade-off in performance for SAR ship detection on the RSDD-SAR dataset.

Summary by alphaXiv.