Abstract
Detecting oriented objects in aerial imagery is critical for remote sensing but faces significant challenges due to domain shift—variations in sensor types, lighting, and geography. This paper introduces GOOD (Generalized Oriented Object Detector), a framework that leverages CLIP-driven style hallucination to train robust detectors capable of generalizing to unseen environments. By enforcing rotation-aware and style consistency, GOOD addresses the twin problems of degraded category recognition and imprecise orientation prediction.
Introduction
Oriented Object Detection (OOD) extends standard object detection by predicting a rotation angle for each bounding box, which is essential for objects like ships and aircraft in satellite imagery. However, models trained on one dataset often fail when deployed in different visual conditions. A "ship" might look distinct under various atmospheric conditions or sensor resolutions, leading to a "domain shift" that cripples performance.
The Problem: Domain Shift in OOD
Domain shift in aerial imagery manifests in two primary ways:
- Degraded Category Prediction: The model fails to identify the object class entirely due to unfamiliar visual features.
- Imprecise Orientation: Even if the object is detected, the predicted angle is often inaccurate. This occurs because the geometric features used for rotation estimation are highly sensitive to style changes.
The GOOD Framework
The authors propose three core components to mitigate these issues:
1. CLIP-driven Style Hallucination
Standard data augmentations are often insufficient for the vast diversity of aerial imagery. GOOD utilizes the pre-trained CLIP (Contrastive Language-Image Pre-training) model to "hallucinate" diverse styles. By extracting style statistics from CLIP's feature space and applying them to the detector's features via Adaptive Instance Normalization (AdaIN), the model learns to recognize objects across a simulated spectrum of environments.
2. Rotation-aware Consistency (RAC)
To stabilize orientation prediction, the RAC objective ensures that an object's features remain consistent across different hallucinated styles. This applies to both the initial horizontal proposals and the final rotated regions, forcing the model to learn style-invariant geometric representations.
3. Style Consistency (SEC)
Using Jensen-Shannon Divergence (JSD), the SEC module minimizes the discrepancy in category predictions between original and style-hallucinated images. This encourages the classifier to focus on the intrinsic content of the object rather than the background or sensor-specific style.
Experimental Results
The GOOD framework was evaluated using a source-to-target protocol across major datasets including FAIR1M, DOTA, SODA, and HRSC.
- Performance Gains: GOOD achieved significant improvements in mean Average Precision (mAP), outperforming baselines like ReDet and RoI-Transformer by up to 3.29% on DOTA and 3.08% on SODA.
- Geometric Robustness: Quantitative analysis showed a marked reduction in Root-Mean-Square Error (RMSE) for rotation angles, confirming that the model's orientation predictions are more stable under domain shift.
- Zero-Shot Generalization: The CLIP-driven hallucination effectively prepared the model for target domains it never encountered during training, bridging the gap between synthetic and real-world performance.
Conclusion and Future Work
GOOD represents a significant step toward robust, real-world aerial object detection. By treating style as a variable to be mastered during training rather than a nuisance at test time, it achieves state-of-the-art generalization. Future research will focus on adapting these techniques to one-stage detectors and exploring "Open-World" scenarios where models must also handle novel, unseen object categories.
Summary updated