Source-linked AI summary
Oriented R-CNN for Object Detection
Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, Junwei Han
TL;DR
Existing two-stage oriented detectors are limited by time-consuming proposal generation. Oriented R-CNN introduces a lightweight oriented RPN and oriented detection head, achieving strong benchmark accuracy with competitive efficiency.
Problem
Proposal generation is the computational bottleneck in proposal-based oriented detectors because existing schemes are expensive.
Method
Oriented R-CNN uses an oriented RPN to directly generate oriented proposals, followed by rotated RoI alignment and an oriented head for classification and regression.
Results
75.87% mAP on DOTA and 96.50% mAP on HRSC2016 were achieved at 15.1 FPS for 1024×1024 images on a single RTX 2080Ti.
Takeaways & Limitations
The authors present oriented R-CNN as a practical baseline combining competitive two-stage accuracy with efficiency comparable to one-stage oriented detectors.
Abstract
from arXiv · showhide
Current state-of-the-art two-stage detectors generate oriented proposals through time-consuming schemes. This diminishes the detectors' speed, thereby becoming the computational bottleneck in advanced oriented object detection systems. This work proposes an effective and simple oriented object detection framework, termed Oriented R-CNN, which is a general two-stage oriented detector with promising accuracy and efficiency. To be specific, in the first stage, we propose an oriented Region Proposal Network (oriented RPN) that directly generates high-quality oriented proposals in a nearly cost-free manner. The second stage is oriented R-CNN head for refining oriented Regions of Interest (oriented RoIs) and recognizing them. Without tricks, oriented R-CNN with ResNet50 achieves state-of-the-art detection accuracy on two commonly-used datasets for oriented object detection including DOTA (75.87% mAP) and HRSC2016 (96.50% mAP), while having a speed of 15.1 FPS with the image size of 1024$\times$1024 on a single RTX 2080Ti. We hope our work could inspire rethinking the design of oriented detectors and serve as a baseline for oriented object detection. Code is available at https://github.com/jbwang1997/OBBDetection.
1. Introduction
Existing proposal-based oriented detectors are slowed by computationally expensive oriented-proposal generation. Oriented R-CNN addresses this bottleneck with a lightweight oriented RPN and achieves strong accuracy and efficiency.
- Proposal-based oriented detectors first generate oriented proposals, then refine and classify them, but proposal generation is computationally expensive.
- Rotated RPN uses 54 anchors per location, while RoI Transformer reduces anchors through a complex process with expensive computation.The rotated RPN uses 3 scales × 3 ratios × 6 angles.
- The paper proposes oriented R-CNN, a simple two-stage detector designed to generate high-quality oriented proposals directly while maintaining competitive efficiency.
- The oriented RPN is a lightweight fully convolutional network using six regression outputs and midpoint offset representation, followed by rotated RoI alignment, classification, and regression.Its parameter count is much smaller than rotated RPN and RoI Transformer+ implementations.
- 75.87% mAP on DOTA and 96.50% mAP on HRSC2016 were achieved at 15.1 FPS for 1024×1024 images on one RTX 2080Ti.
2. Related Work
Oriented object detection addresses localization problems caused by horizontal boxes, while prior proposal-based and one-stage methods make different efficiency and feature-alignment trade-offs. Oriented R-CNN remains proposal-based but targets efficient oriented-proposal generation.
- Oriented detection extends object detection to objects whose horizontal boxes may include excess background or multiple objects.
- RoI Transformer learns rotated RoIs from horizontal RoIs, improving oriented detection accuracy but increasing network complexity through fully connected layers and RoI alignment.
- One-stage and anchor-free methods output classes and oriented boxes without region proposal generation or RoI alignment.
- Figure 2 depicts oriented R-CNN as a two-stage FPN detector whose oriented RPN generates proposals before the oriented R-CNN head classifies and regresses them.
- Oriented R-CNN stays within proposal-based detection while focusing on a high-efficiency oriented RPN to address the proposal-generation bottleneck.
3. Oriented R-CNN
Oriented R-CNN combines a lightweight oriented RPN with a second-stage head for classifying and refining oriented proposals. Its midpoint offset representation parameterizes oriented boxes for proposal generation and regression.
- The detector has two stages: an oriented RPN generates proposals, while the oriented R-CNN head performs proposal classification and regression.
- Oriented RPN: The oriented RPN uses a fully convolutional design with six regression outputs per anchor to directly decode oriented proposals.The six offsets are δx, δy, δw, δh, δα, and δβ.
- Midpoint Offset Representation: Midpoint offset representation encodes an oriented box with an external rectangle and offsets from the top and right side midpoints.The representation uses six parameters (x, y, w, h, ∆α, ∆β) and derives the four box vertices.
- Midpoint Offset Representation: The representation enables regression by predicting external-rectangle parameters and inferring the two midpoint offsets.
- Loss Function: Anchor labels use IoU thresholds of 0.7 and 0.3, with intermediate anchors ignored during training.An anchor is positive above 0.7 or when it has the highest overlap above 0.3; anchors below 0.3 are negative.
- Loss Function: The training loss combines classification and regression terms, with cross entropy for objectness and Smooth L1 for six-dimensional proposal offsets.
4. Experiments
Experiments evaluate oriented R-CNN on DOTA and HRSC2016, measuring proposal recall, detection accuracy, and speed. The method achieves strong recall and state-of-the-art accuracy while maintaining comparable inference efficiency.
- Datasets: Experiments use the DOTA and HRSC2016 oriented object detection benchmarks.DOTA contains 2,806 images and 188,282 instances across 15 classes; HRSC2016 targets arbitrary-oriented ship detection.
- Evaluation of Oriented RPN: 92.80% recall is achieved by oriented RPN with 2,000 proposals on the DOTA validation set.Reducing proposals from 2,000 to 1,000 decreases recall by only 0.6%, whereas 300 proposals cause a sharp drop.
- Evaluation of Oriented RPN: 1,000 proposals are selected at test time for both datasets to trade off inference speed and detection accuracy.The generated proposals localize objects across sizes, aspect ratios, directions, and densities.
- Comparison with State-of-the-Arts: 75.87% mAP is obtained on DOTA with an R-50-FPN backbone, surpassing all comparison methods.R-101-FPN reaches 76.28% mAP, while multi-scale training and testing with R-50-FPN reaches 80.87% mAP.
- Speed versus Accuracy: 15.1 FPS is achieved at 75.87% mAP on 1024×1024 inputs using a single RTX 2080Ti.Under the same R-50-FPN setting, the speed is comparable to other methods and close to one-stage detectors, while accuracy is higher than one-stage detectors.
5. Conclusions
The paper concludes that oriented R-CNN is a practical two-stage detector for arbitrary-oriented object detection. Experiments on two challenging benchmarks show competitive accuracy and efficiency relative to advanced two-stage and one-stage oriented detectors.
- 5. Conclusions: Oriented R-CNN is a practical two-stage detector for arbitrary-oriented object detection in images.The conclusion frames the method as applicable to the paper’s target detection setting.
- 5. Conclusions: Experiments on two challenging oriented object detection benchmarks support competitive accuracy and efficiency.The conclusion compares accuracy with advanced two-stage detectors and efficiency with one-stage oriented detectors.