open-mmlab/mmdetection ↗

About open-mmlab/mmdetection

MMDetection is an open-source object detection toolbox built on PyTorch and part of the OpenMMLab ecosystem. It provides a modular, component-based framework that allows researchers and practitioners to construct customized detection systems by combining different architectural elements. The toolbox supports multiple detection tasks including object detection, instance segmentation, panoptic segmentation, and semi-supervised object detection, with implementations spanning decades of research including foundational methods like Fast R-CNN and modern approaches like DETR, RTMDet, and MM-Grounding-DINO.

The framework emphasizes efficiency and state-of-the-art performance, with all basic bounding box and mask operations implemented on GPU. It includes a comprehensive model zoo containing dozens of detector architectures alongside various backbone networks (ResNet, Vision Transformers, ConvNeXt, Swin Transformers), neck designs, and loss functions drawn from recent literature. Notable achievements include winning the COCO Detection Challenge in 2018 and recent advances with RTMDet, which obtains state-of-the-art results on real-time instance segmentation and rotated object detection while maintaining excellent parameter-accuracy trade-offs.

The project caters to a broad audience from academic researchers to industry practitioners, offering extensive documentation, configuration-based training workflows, dataset preparation utilities, and tools for fine-tuning, testing, and model submission. As part of the larger OpenMMLab ecosystem, MMDetection integrates with complementary libraries like MMEngine for training infrastructure and MMCV for general computer vision functionality, establishing it as a foundational resource for detection research and applications.