Source-linked AI summary
MMRotate: A Rotated Object Detection Benchmark using PyTorch
Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, Wenwei Zhang, Kai Chen
TL;DR
Rotated object detection needs representations and implementations that handle diverse angle conventions and algorithmic components. MMRotate provides an open-source PyTorch framework integrating 18 methods, benchmarks, trained models, and tools; it supports three common angle definitions and reports component-dependent performance patterns. The toolbox is intended to facilitate research and industrial deployment across diverse domains.
Problem
Rotated object detection uses diverse angle definitions, optimization strategies, operators, and algorithm implementations, making unified comparison and reuse important.
Method
MMRotate provides an open-source PyTorch framework integrating rotated detection algorithms, angle-conversion support, trained models, benchmarks, and utility tools.
Results
MMRotate implements 18 rotated object detection methods and supports OpenCV, long edge 90°, and long edge 135° angle definitions.
Takeaways & Limitations
The unified benchmark is designed to enable fair algorithm comparisons, investigation of effective components, and deployment across academic and industrial domains.
Abstract
from arXiv · showhide
We present an open-source toolbox, named MMRotate, which provides a coherent algorithm framework of training, inferring, and evaluation for the popular rotated object detection algorithm based on deep learning. MMRotate implements 18 state-of-the-art algorithms and supports the three most frequently used angle definition methods. To facilitate future research and industrial applications of rotated object detection-related problems, we also provide a large number of trained models and detailed benchmarks to give insights into the performance of rotated object detection. MMRotate is publicly released at https://github.com/open-mmlab/mmrotate.
1 INTRODUCTION
Rotated object detection uses oriented boxes to better align objects than horizontal boxes, and MMRotate unifies diverse methods and components for research and applications.
- Oriented bounding boxes better align objects than horizontal boxes for more accurate identification.
- MMRotate provides a unified framework for training, inference, and evaluation of rotated object detection approaches.
- The toolbox implements 18 rotated object detection methods, 10 CUDA speed-up operators, and 12 losses.
- Integrating algorithms improves code reusability and simplifies implementation while enabling fair comparisons and investigation of effective components.
- MMRotate includes documentation, dataset scripts, API documentation, a model zoo, tutorials, a user manual, and trained models and benchmarks.
2 RELATED WORK
Related work spans rotated detection representations, feature and label-assignment methods, and open-source toolboxes; MMRotate addresses limitations in existing implementations while broadening algorithm coverage.
- Rotated detection uses angle-based box representations and can also represent objects with two-dimensional Gaussian distributions or point sets.
- Feature alignment methods refine feature maps, align convolutional features, or produce rotation-equivariant RoI features for oriented objects.
- Label-assignment methods replace or supplement IoU-based selection with matching degrees, posterior-IoU sampling, or shape-adaptive strategies.
- Existing open-source toolboxes vary in framework and detector coverage, while some rely on outdated MMDetection versions or are no longer maintained.
3 ROTATED OBJECT DETECTION STUDIES
MMRotate studies how angle definitions, backbones, losses, and training tools affect rotated detector performance, resource use, and flexibility.
- The study exchanges angle definitions, backbones, and losses across approaches to measure performance, memory usage, and inference time.
- OpenCV, long edge 90°, and long edge 135° definitions are switchable through configuration, with an API for other angle conventions.
- Swin-T significantly outperforms ResNet50 in RoI Trans, but inference is 21% slower.
- KLD achieves the best mAP with RetinaNet, whereas KFIoU achieves the best mAP with R3Det under the OpenCV definition.
- Mixed-precision fp16 training produces similar mAP to the original model, while utilities support visualization, confusion-matrix analysis, and huge-image inference.
4 CONCLUSIONS
MMRotate is a flexible PyTorch benchmark for rotated object detection intended to support deployment across industrial and academic settings.
- MMRotate is released under the Apache-2.0 license as a flexible, easy-to-use PyTorch benchmark for visual object rotation detection.
- The benchmark is designed to facilitate deployment in diverse industrial applications and academic research.
- The authors plan to continue improving the benchmark and supporting representative detection methods with community participation.