33K ·GitHub
What is this?
MMDetection is an open source object detection toolbox built on PyTorch, part of the OpenMMLab project, that supports object detection, instance segmentation, panoptic segmentation, and semi-supervised detection.
MMDetection is an open source object detection toolbox built on PyTorch, part of the OpenMMLab project, that supports object detection, instance segmentation, panoptic segmentation, and semi-supervised detection.Is this a framework I build on, or a ready-to-run app?It is a modular research framework/library: it decomposes the detection pipeline into components (backbones, necks, heads) so you can combine modules to build custom detectors.
It is a modular research framework/library: it decomposes the detection pipeline into components (backbones, necks, heads) so you can combine modules to build custom detectors.How do I get it running?Follow the official Installation guide; after installing, use the User Guides on training/testing, running inference with existing models, and preparing datasets.
Follow the official Installation guide; after installing, use the User Guides on training/testing, running inference with existing models, and preparing datasets.What do I need installed before this works?It requires PyTorch 1.8+ and relies on two sibling OpenMMLab libraries, MMEngine (training) and MMCV (core CV ops), which it depends on heavily.
It requires PyTorch 1.8+ and relies on two sibling OpenMMLab libraries, MMEngine (training) and MMCV (core CV ops), which it depends on heavily.Is this still the actively maintained version, or should I look at something newer?The project is actively updated with new releases (e.g., v3.3.0 and MM-Grounding-DINO in 2024), but users migrating from the older 2.x series need to follow a separate migration guide since APIs changed.
The project is actively updated with new releases (e.g., v3.3.0 and MM-Grounding-DINO in 2024), but users migrating from the older 2.x series need to follow a separate migration guide since APIs changed.How big is the model zoo?Huge — it covers dozens of architectures including Faster R-CNN, Mask R-CNN, RetinaNet, DETR variants, Swin Transformer, YOLOX, RTMDet, Grounding DINO, and many more across detection, instance and panoptic segmentation.
Huge — it covers dozens of architectures including Faster R-CNN, Mask R-CNN, RetinaNet, DETR variants, Swin Transformer, YOLOX, RTMDet, Grounding DINO, and many more across detection, instance and panoptic segmentation.What is RTMDet and why is it highlighted?RTMDet is a family of real-time, fully convolutional single-stage detectors released by the project that achieves state-of-the-art parameter-accuracy trade-offs on object detection and also leads in instance and rotated object detection, e.g. 52.8 AP at 322 FPS on COCO.
RTMDet is a family of real-time, fully convolutional single-stage detectors released by the project that achieves state-of-the-art parameter-accuracy trade-offs on object detection and also leads in instance and rotated object detection, e.g. 52.8 AP at 322 FPS on COCO.What is MM-Grounding-DINO?It's the team's open-source reproduction and improvement of Grounding DINO, a model unifying open-vocabulary object detection and phrase grounding, with released Swin-B and Swin-L pretrained weights and evaluations across OOD, REC, phrase grounding, OVD, and fine-tuning settings.
It's the team's open-source reproduction and improvement of Grounding DINO, a model unifying open-vocabulary object detection and phrase grounding, with released Swin-B and Swin-L pretrained weights and evaluations across OOD, REC, phrase grounding, OVD, and fine-tuning settings.How does MMDetection compare in training speed to other toolboxes?The authors state its training speed is faster than or comparable to other codebases like Detectron2, maskrcnn-benchmark, and SimpleDet.
The authors state its training speed is faster than or comparable to other codebases like Detectron2, maskrcnn-benchmark, and SimpleDet.What are similar projects?deep-learning-for-image-processing, Swin-Transformer, PaddleDetection, GroundingDINO, detr, VisionHOPE, yolov10, ultralytics
deep-learning-for-image-processing, Swin-Transformer, PaddleDetection, GroundingDINO, detr, VisionHOPE, yolov10, ultralyticsWhich topics is it in?swin-transformer, mask-rcnn, instance-segmentation, vision-transformer, yolo, segmentation, object-detection, transformer, computer-vision, pytorch, deep-learning, research
swin-transformer, mask-rcnn, instance-segmentation, vision-transformer, yolo, segmentation, object-detection, transformer, computer-vision, pytorch, deep-learning, research