RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark

TL;DR

This work addresses angle boundary discontinuity in rotated object detection by reframing angle encodings as dimensional mappings and enforcing a unit-circle constraint via the Unit Circle Resolver (UCR). The authors show that existing resolvers neglect unit-circle constraints, causing biases, and they formulate a loss $L_{uc}$ to constrain encodings; applying UCR to H2RBox-v2 improves angle prediction and enables accurate pseudo-rotated labeling. Using these pseudo-labels, they build RSAR, the largest multi-class rotated SAR object detection dataset to date, and demonstrate that UCR enhances angle accuracy and detection performance across optical and SAR benchmarks—sometimes surpassing fully supervised methods. The RSAR dataset, along with code, provides a strong foundation for advancing rotated-SAR detection and weakly supervised learning in this domain.

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.

Figures (6)

• Figure 1: A comparison between prediction results from a weakly supervised model and the ground truth (GT). The weakly supervised model's accuracy in predicting the object angle requires further improvement.
• Figure 2: Diagram analysis of the angle boundary discontinuity problem from the unified perspective of dimensional mapping. (a) One-dimensional space leads to the problem of angle boundary discontinuity. (b) Mapping one-dimensional values to two-dimensional and three-dimensional spaces helps address the issue. (c) Existing methods overlook the unit circle constraint inherent in the angle encoding states, leading to a many-to-one problem that introduces biases in model optimization and predictions. Our unified perspective clarifies the angle boundary discontinuity issue and exposes the potential shortcomings of existing methods.
• Figure 3: Visualization of images from RSAR and SARDet-100K. Rotated annotations in RSAR offer higher location accuracy compared to horizontal annotations in SARDet-100K. Rotated SAR object detection presents greater challenges than horizontal SAR object detection.
• Figure 4: Statistical visualization of attributes for each category in RSAR. (a) The angle distribution of instances for each category (expressed in $le_{90}$ angle notation). (b) The aspect ratio distribution of instances for each category. (c) The percentage of instances for each category. (d) The average instance pixel area for each category.
• Figure 5: Comparison of visualized results on RSAR and DOTA-v1.0 in two-dimensional mapping. AE represents angle encoding, indicating the $(\cos\theta, \ \sin\theta)$ of the model's prediction, while AE$^2$ denotes their sum of squares (i.e., $AE^2=\sin^2\theta+\cos^2\theta$). We obtain the predicted values for all angle encodings of the bounding boxes on the test set and display their probability distribution statistics in the image on the left. The white regions in the probability distribution diagram correspond to areas where angle encoding has a higher probability of occurrence. Due to the unrestricted distribution in the angle encoding state space of the no-limit method, angle predictions may lack accuracy in certain scenarios. Our method significantly enhances the accuracy of angle prediction in the model.