[Paper Review] ReDet: A Rotation-equivariant Detector for Aerial Object Detection

Jiaming Han, Jian Ding+2 more|arXiv (Cornell University)|Mar 13, 2021

Advanced Neural Network Applications45 references39 citations

TL;DR

ReDet introduces a rotation-equivariant backbone and a Rotation-invariant RoI Align to achieve fully rotation-invariant features for aerial object detection, achieving state-of-the-art mAP while reducing model size.

ABSTRACT

Recently, object detection in aerial images has gained much attention in computer vision. Different from objects in natural images, aerial objects are often distributed with arbitrary orientation. Therefore, the detector requires more parameters to encode the orientation information, which are often highly redundant and inefficient. Moreover, as ordinary CNNs do not explicitly model the orientation variation, large amounts of rotation augmented data is needed to train an accurate object detector. In this paper, we propose a Rotation-equivariant Detector (ReDet) to address these issues, which explicitly encodes rotation equivariance and rotation invariance. More precisely, we incorporate rotation-equivariant networks into the detector to extract rotation-equivariant features, which can accurately predict the orientation and lead to a huge reduction of model size. Based on the rotation-equivariant features, we also present Rotation-invariant RoI Align (RiRoI Align), which adaptively extracts rotation-invariant features from equivariant features according to the orientation of RoI. Extensive experiments on several challenging aerial image datasets DOTA-v1.0, DOTA-v1.5 and HRSC2016, show that our method can achieve state-of-the-art performance on the task of aerial object detection. Compared with previous best results, our ReDet gains 1.2, 3.5 and 2.6 mAP on DOTA-v1.0, DOTA-v1.5 and HRSC2016 respectively while reducing the number of parameters by 60\% (313 Mb vs. 121 Mb). The code is available at: \url{https://github.com/csuhan/ReDet}.

Motivation & Objective

• Motivate and address arbitrary object orientation in aerial imagery.
• Incorporate rotation-equivariant networks into the detector backbone.
• Develop a Rotation-invariant RoI Align to produce fully rotation-invariant RoI features.
• Demonstrate state-of-the-art performance on DOTA-v1.0, DOTA-v1.5, and HRSC2016.
• Show model-size and accuracy improvements over baselines.

Proposed method

• Adopt a rotation-equivariant backbone (ReResNet) with ReFPN to produce rotation-equivariant feature maps across N orientation channels.
• Introduce RiRoI Align that performs spatial RoI warping plus orientation-channel switching and interpolation to yield rotation-invariant RoI features.
• Use an RoI Transformer to generate rotated RoIs and apply RiRoI Align for RoI-wise classification and bbox regression.
• Train and fine-tune with standard detection pipelines (RPN, RoIHead) on aerial datasets with oriented bounding boxes.
• Demonstrate parameter efficiency via shared weights and reduced parameter count due to rotation-equivariant design.

Figure 1: Illustration of our method (top) and comparisons of RRoI warping (bottom) . CNN features are not equivariant to the rotation $T_{r}$ , i.e. , feeding a rotated image to CNNs is not the same as rotating feature maps of the original image. Therefore, the corresponding RoI features are not in

Experimental results

Research questions

• RQ1 Can rotation-equivariant backbones reduce the need for large orientation-specific parameters in aerial object detectors?
• RQ2 Does a Rotation-invariant RoI Align effectively extract orientation-invariant features from rotation-equivariant backbones?
• RQ3 What is the impact of ReDet on key aerial detection benchmarks (DOTA-v1.0, DOTA-v1.5, HRSC2016) compared to strong baselines?
• RQ4 How does ReDet influence model size and accuracy trade-offs compared to non-rotational augmentation methods?

Key findings

• ReDet achieves 80.10 mAP on DOTA-v1.0, 76.80 mAP on DOTA-v1.5, and 90.46 mAP on HRSC2016, outperforming prior best results by 1.2, 3.5, and 2.6 mAP respectively.
• ReDet reduces model size by about 60% (313 Mb baseline vs. 121 Mb) while delivering competitive to superior performance.
• RiRoI Align outperforms traditional RRoI Align, with l=2 orientation interpolation yielding the best mAP gains (66.86 mAP in ablations).
• Rotation-equivariant backbones (ReResNet+ReFPN) improve detection performance with significantly fewer parameters, especially under C8 rotation groups.
• Compared with rotation augmentation baselines under similar schedules, ReDet shows substantial mAP gains with comparable training time, and better parameter efficiency.

Figure 2: Model size vs. accuracy (mAP) on DOTA-v1.5. We evaluate RetinaNet OBB [ 18 ] , Faster R-CNN OBB (FR) [ 27 ] , Mask R-CNN (Mask) [ 11 ] and Hybrid Task Cascade (HTC) [ 2 ] with ResNet18 (R18) and ResNet50 (R50) backbones. Note all algorithms are our re-implemented version for DOTA, which is