Introduction
Aerial object detection presents a unique challenge: unlike terrestrial images where objects often follow a standard "upright" orientation due to gravity, aerial imagery captures objects at arbitrary angles. Standard Convolutional Neural Networks (CNNs) are translation-equivariant but lack rotation-equivariance. Consequently, they often require massive data augmentation or excessive parameters to "brute-force" learn every possible orientation, leading to redundant features and unstable detections. ReDet (Rotation-equivariant Detector) addresses this by explicitly modeling the physics of rotation within the network architecture.
Proposed Method: ReDet
ReDet introduces a two-pronged approach to achieve rotation-invariant detection with high efficiency.
1. Rotation-equivariant Backbone (ReResNet)
Instead of standard convolutions, ReDet utilizes group convolutions (based on the $e2cnn$ framework). This ensures that when an input image rotates, the feature maps undergo a predictable transformation—rotating spatially and shifting across "orientation channels." This mathematical consistency allows the network to share weights across different angles, significantly reducing parameter count.
2. Rotation-invariant RoI Align (RiRoI Align)
While the backbone is equivariant, the final classification head requires rotation-invariant features—a "plane" should be recognized as such regardless of its heading. Standard Rotated RoI (RRoI) Align only accounts for spatial warping. RiRoI Align extends this by performing both spatial alignment and orientation alignment, circularly shifting orientation channels to match the object's predicted heading.
Experimental Evaluation
The authors evaluated ReDet on three major aerial benchmarks: DOTA-v1.0, DOTA-v1.5, and HRSC2016. These datasets contain hundreds of thousands of instances across diverse categories, including challenging long, thin objects like ships and planes.
Key Results and Efficiency
ReDet establishes new state-of-the-art performance while maintaining a remarkably small footprint. It achieved 80.10 mAP on DOTA-v1.0 and 76.80 mAP on DOTA-v1.5, outperforming previous leading models.
Crucially, ReDet reduced the parameter count from 313 MB in previous top-performing models to just 121 MB—a 60% reduction. On the HRSC2016 ship dataset, it reached 90.46 mAP, demonstrating its robustness for objects with extreme aspect ratios and orientations.
Discussion and Future Work
The core innovation of ReDet lies in the strategic split between equivariance in the backbone and invariance in the detection head. By treating different rotations as symmetric patterns rather than unique features, the model avoids wasting parameters.
Current limitations include the use of discrete rotation groups (e.g., 8 or 16 angles), which approximates but does not perfectly achieve continuous rotation equivariance. Additionally, group convolutions introduce a slight training overhead (approx. 18% longer). Future research may explore more complex symmetry groups and extend these principles to tasks like instance segmentation in aerial scenes.
Summary updated