RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (2501.04440v1)

Abstract: Rotated object detection has made significant progress in the optical remote sensing. However, advancements in the Synthetic Aperture Radar (SAR) field are laggard behind, primarily due to the absence of a large-scale dataset. Annotating such a dataset is inefficient and costly. A promising solution is to employ a weakly supervised model (e.g., trained with available horizontal boxes only) to generate pseudo-rotated boxes for reference before manual calibration. Unfortunately, the existing weakly supervised models exhibit limited accuracy in predicting the object's angle. Previous works attempt to enhance angle prediction by using angle resolvers that decouple angles into cosine and sine encodings. In this work, we first reevaluate these resolvers from a unified perspective of dimension mapping and expose that they share the same shortcomings: these methods overlook the unit cycle constraint inherent in these encodings, easily leading to prediction biases. To address this issue, we propose the Unit Cycle Resolver, which incorporates a unit circle constraint loss to improve angle prediction accuracy. Our approach can effectively improve the performance of existing state-of-the-art weakly supervised methods and even surpasses fully supervised models on existing optical benchmarks (i.e., DOTA-v1.0 dataset). With the aid of UCR, we further annotate and introduce RSAR, the largest multi-class rotated SAR object detection dataset to date. Extensive experiments on both RSAR and optical datasets demonstrate that our UCR enhances angle prediction accuracy. Our dataset and code can be found at: https://github.com/zhasion/RSAR.

• The paper introduces the Unit Cycle Resolver (UCR) to improve angle prediction in weakly supervised rotated object detection by enforcing a unit cycle constraint and presents RSAR, a large new rotated SAR dataset.
• The Unit Cycle Resolver significantly enhances angle prediction accuracy on both SAR and optical datasets like DOTA, demonstrating its effectiveness beyond SAR imagery.
• The UCR can be integrated into existing object detection pipelines to enhance angle predictions, while the extensive RSAR dataset provides a valuable resource for future SAR research.

An Expert Review of "RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark"

The paper "RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark" explores the domain of rotated object detection, particularly within the context of Synthetic Aperture Radar (SAR) imagery. As rotated object detection techniques become increasingly precise and applicable, this research addresses the lag in SAR advancements by introducing innovative methodologies both at the algorithmic level and in dataset creation.

The authors introduce the concept of the Unit Cycle Resolver (UCR) to address the angle prediction challenges faced by existing weakly supervised models. They identify a critical limitation in existing angle resolvers: the oversight of the unit cycle constraint intrinsic to encoding formulations. This oversight often results in prediction biases that degrade model performance. The UCR enhances prediction accuracy by incorporating a unit circle constraint loss, which ensures that angle resolutions remain true to their inherent geometrical constraints, aligning more closely with the actual physical orientation of rotated objects.

A key contribution of the paper is not merely theoretical but also practical. The authors apply the UCR in annotating a new, extensive dataset: the RSAR dataset. RSAR stands as the largest multi-class rotated SAR object detection dataset available, comprising 95,842 images and 183,534 annotated instances across six categories. The dataset mitigates the annotation inefficiencies and costs typically associated with SAR data by leveraging weakly supervised models to generate pseudo-rotated boxes, which are then refined manually, enhancing annotation efficiency significantly.

The experimental results demonstrate that the UCR can markedly improve angle prediction accuracy across both SAR and optical datasets. Evaluations on the DOTA-v1.0 dataset illustrate that the UCR-enhanced models surpass the performance of many fully supervised counterparts on standard optical benchmarks. Such performance gains not only validate the theoretical insights provided by the unified perspective on dimensional mapping but substantiate the potential of UCR to set new baselines in SAR rotated object detection.

The implications of this work are manifold. Practically, the UCR can be integrated into existing object detection pipelines to enhance angle-related predictions without the necessity of full re-annotation from scratch. Theoretically, it offers a refined perspective on how encoding states should be constrained to reflect precise angle predictions, potentially influencing future research directions on handling periodic ambiguities and boundary discontinuities in varied domains.

In summary, this paper makes a significant contribution to the field by providing both a robust methodological advancement in the form of the Unit Cycle Resolver and a substantial dataset, RSAR, which promises to catalyze future research endeavors in SAR object detection. Future research could explore further enhancements to UCR, its integration with other forms of remote sensing data, and its application to fully supervised models to push the boundaries of both SAR and optical detection capabilities.