MMDetection is an open-source object detection and instance segmentation toolbox developed by the MMLab team. It provides a comprehensive set of detection methods, modules, and components, built upon the codebase that won the COCO Detection Challenge 2018. The toolbox supports a wide range of popular detection frameworks, including single-stage, two-stage, and multi-stage methods, as well as various general modules and techniques. It includes training and inference code, over 200 pre-trained models, and is designed to be flexible and efficient, with high performance and compatibility with multiple deep learning frameworks. The toolbox is built with PyTorch and features a modular design, allowing users to customize detection frameworks by combining different modules. It supports multiple frameworks out of the box, and is optimized for performance, with training speed comparable to other leading codebases. MMDetection also includes a benchmarking system that evaluates different methods, components, and hyperparameters, enabling researchers to compare results across different approaches. The toolbox supports a variety of detection methods, including SSD, RetinaNet, Faster R-CNN, Mask R-CNN, and FCOS, among others. It also includes advanced modules such as mixed precision training, Soft NMS, OHEM, DCN, and more. The architecture of MMDetection is modular, with components such as backbone, neck, dense head, ROI extractor, and ROI head, which can be combined to create custom detection frameworks. The toolbox is evaluated on the COCO 2017 dataset, with results showing that it achieves high performance and efficiency. It supports various training scales and hyperparameters, and includes extensive studies on regression losses, normalization layers, and training scales. The toolbox is continuously developed and updated, with a flexible and efficient training pipeline that supports distributed training across multiple GPUs. MMDetection provides a flexible and efficient toolkit for researchers to reimplement existing methods and develop new detectors, and is available at https://github.com/open-mmlab/mmdetection. It is designed to serve the growing research community by providing a comprehensive and well-documented platform for object detection and instance segmentation research.MMDetection is an open-source object detection and instance segmentation toolbox developed by the MMLab team. It provides a comprehensive set of detection methods, modules, and components, built upon the codebase that won the COCO Detection Challenge 2018. The toolbox supports a wide range of popular detection frameworks, including single-stage, two-stage, and multi-stage methods, as well as various general modules and techniques. It includes training and inference code, over 200 pre-trained models, and is designed to be flexible and efficient, with high performance and compatibility with multiple deep learning frameworks. The toolbox is built with PyTorch and features a modular design, allowing users to customize detection frameworks by combining different modules. It supports multiple frameworks out of the box, and is optimized for performance, with training speed comparable to other leading codebases. MMDetection also includes a benchmarking system that evaluates different methods, components, and hyperparameters, enabling researchers to compare results across different approaches. The toolbox supports a variety of detection methods, including SSD, RetinaNet, Faster R-CNN, Mask R-CNN, and FCOS, among others. It also includes advanced modules such as mixed precision training, Soft NMS, OHEM, DCN, and more. The architecture of MMDetection is modular, with components such as backbone, neck, dense head, ROI extractor, and ROI head, which can be combined to create custom detection frameworks. The toolbox is evaluated on the COCO 2017 dataset, with results showing that it achieves high performance and efficiency. It supports various training scales and hyperparameters, and includes extensive studies on regression losses, normalization layers, and training scales. The toolbox is continuously developed and updated, with a flexible and efficient training pipeline that supports distributed training across multiple GPUs. MMDetection provides a flexible and efficient toolkit for researchers to reimplement existing methods and develop new detectors, and is available at https://github.com/open-mmlab/mmdetection. It is designed to serve the growing research community by providing a comprehensive and well-documented platform for object detection and instance segmentation research.