Fourier Angle Alignment for Oriented Object Detection in Remote Sensing
Fourier Angle Alignment (FAA) improves rotated object detection accuracy in remote sensing, achieving 78.72% mAP on DOTA-v1.0 with frequency-based orientation estimation.
Key Findings
Methodology
This work leverages Fourier transforms to analyze object orientation in the frequency domain, proposing FAA which estimates dominant angles and aligns features at multiple stages. FAAFusion adjusts high-level features based on low-level frequency estimates within FPN, ensuring scale consistency. FAA Head pre-aligns RoI features to a canonical orientation, decoupling classification and regression tasks. The framework integrates seamlessly into existing detectors, enhancing robustness against rotation and scale variations.
Key Results
- On DOTA-v1.0, the method achieves 78.72% mAP, surpassing previous SOTA by 0.68%, validating the frequency domain orientation estimation's effectiveness.
- On DOTA-v1.5, it reaches 72.28% mAP, especially improving detection of small and complex targets.
- On HRSC2016, accuracy in angle prediction and overall detection performance improved by 1.23%, demonstrating broad applicability.
Significance
This approach addresses longstanding issues in rotated object detection by providing a robust, frequency-based orientation estimation method. It effectively resolves the conflict between scale-invariant classification and rotation-sensitive regression, leading to significant performance gains. The framework offers a new theoretical perspective and practical tools for remote sensing applications, with potential impacts on aerial surveillance, maritime monitoring, and urban planning.
Technical Contribution
The paper introduces a novel frequency-domain orientation estimation mechanism integrated into feature fusion and detection heads. By combining Fourier spectral analysis with deep learning, it achieves scale and task decoupling, improving detection accuracy. The approach is compatible with various backbone networks and enhances the interpretability and robustness of orientation prediction, opening new avenues for rotation-aware detection.
Novelty
This is the first work to incorporate Fourier spectral analysis for target orientation estimation in remote sensing detection. It innovatively exploits the spectral energy concentration properties of objects, enabling robust, scale-invariant angle prediction. Unlike prior spatial domain methods, this frequency-based approach provides a new theoretical foundation for rotation-aware detection.
Limitations
- Frequency domain analysis incurs higher computational costs, which may limit real-time deployment.
- In scenarios with heavy occlusion or blurred targets, spectral signals may be noisy, reducing angle estimation accuracy.
- The method's robustness under extreme clutter or overlapping objects needs further validation and optimization.
Future Work
Future research will explore multi-modal data fusion (e.g., LiDAR, multispectral) to enhance spectral robustness, optimize spectral feature extraction for real-time applications, and extend the framework to tasks like object tracking and pose estimation, aiming for comprehensive scene understanding.
AI Executive Summary
Remote sensing object detection faces significant challenges due to target orientation variability, especially in aerial imagery where objects like ships, vehicles, and aircraft can appear in arbitrary directions. Traditional detectors often struggle with accurate angle prediction, especially across multiple scales, leading to decreased detection performance. Existing solutions, such as oriented bounding box methods and rotation-sensitive convolutions, have made progress but still face limitations in scale-invariance and task decoupling.
This paper introduces Fourier Angle Alignment (FAA), a novel framework that leverages frequency domain analysis to estimate and align object orientations. FAA employs Fourier transforms to extract spectral energy distributions, identifying the dominant orientation of objects robustly. The framework consists of two key modules: FAAFusion, which aligns high-level features with low-level frequency-estimated directions within the feature pyramid, and FAA Head, which pre-aligns RoI features to a canonical angle before classification and regression. These modules effectively address the conflicting requirements of rotation invariance for classification and rotation sensitivity for angle regression.
Experimental results on DOTA-v1.0, DOTA-v1.5, and HRSC2016 datasets demonstrate substantial improvements over state-of-the-art methods. On DOTA-v1.0, the method achieves 78.72% mAP, surpassing previous bests by over 0.6%. Similar gains are observed on DOTA-v1.5 and HRSC2016, validating the approach's robustness and generality. The frequency-based orientation estimation not only enhances detection accuracy but also offers a new theoretical perspective for rotation-aware detection, opening pathways for future multi-modal and real-time applications.
Despite these advances, challenges remain in computational efficiency and robustness under extreme conditions. Future work will focus on multi-modal fusion, real-time optimization, and extending the framework to broader tasks like pose estimation and multi-object tracking, aiming to realize fully autonomous remote sensing systems with high precision and resilience.
Deep Dive
Abstract
In remote sensing rotated object detection, mainstream methods suffer from two bottlenecks, directional incoherence at detector neck and task conflict at detecting head. Ulitising fourier rotation equivariance, we introduce Fourier Angle Alignment, which analyses angle information through frequency spectrum and aligns the main direction to a certain orientation. Then we propose two plug and play modules : FAAFusion and FAA Head. FAAFusion works at the detector neck, aligning the main direction of higher-level features to the lower-level features and then fusing them. FAA Head serves as a new detection head, which pre-aligns RoI features to a canonical angle and adds them to the original features before classification and regression. Experiments on DOTA-v1.0, DOTA-v1.5 and HRSC2016 show that our method can greatly improve previous work. Particularly, our method achieves new state-of-the-art results of 78.72% mAP on DOTA-v1.0 and 72.28% mAP on DOTA-v1.5 datasets with single scale training and testing, validating the efficacy of our approach in remote sensing object detection. The code is made publicly available at https://github.com/gcy0423/Fourier-Angle-Alignment .