MMDetection: OpenMMLab detection toolbox

What is MMDetection?

MMDetection is an open source object detection toolbox built on PyTorch, part of the OpenMMLab project. It packages object detection, instance segmentation, panoptic segmentation, and semi-supervised detection into one modular framework where backbones, necks, and heads swap through config files rather than code edits. It depends on MMEngine and MMCV for training and core CV operations.

Key features: modular design and multi-task support

• ✓Modular design: the README describes the framework as decomposed into separate components — backbones, necks, heads — that you recombine into a custom detector instead of subclassing a monolith.
• ✓Multi-task support out of the box: object detection, instance segmentation, panoptic segmentation, and semi-supervised object detection all run through the same codebase.
• ✓A wide architecture zoo, spanning Fast R-CNN, Faster R-CNN, RetinaNet, Cascade R-CNN, YOLOv3, DETR, Deformable DETR, DINO, GLIP, and RTMDet, plus backbones like Swin Transformer and ConvNeXt.
• ✓RTMDet, the newer real-time detector family, reports 52.8 AP on COCO object detection at 322 FPS (TRT FP16, batch size 1, RTX 3090) per the README's own benchmark table.
• ✓GPU-resident bbox and mask ops — the README states training speed is faster than or comparable to Detectron2, maskrcnn-benchmark, and SimpleDet.
• ✓MM-Grounding-DINO, an open source replication of Grounding DINO for open-vocabulary detection and phrase grounding, ships with released pretrained Swin-B and Swin-L weights.

Installing MMDetection

The README itself doesn't include install commands — it links out to the official 'Installation' guide on mmdetection.readthedocs.io for the actual pip/conda steps, so exact commands aren't given in the source available here. What is documented directly: MMDetection's main branch needs PyTorch 1.8+, and it depends on two sibling OpenMMLab libraries, MMEngine (training) and MMCV (core CV operations), installed alongside it. Beyond that, the specific install sequence is not clearly documented in the README and you'd follow the linked guide.

Running a detection model with config files

MMDetection's README doesn't walk through a config file or a training command on the page itself — it points to separate 'Learn about Configs' and 'Inference with existing models' guides in the docs, so the exact CLI invocation is not clearly documented in the source available here. Conceptually, the workflow is config-driven: you pick or write a config file that assembles backbone, neck, and head modules, then point the train/test scripts at that config instead of editing Python training code directly. There's a dedicated 'Train with customized datasets' guide, plus a 'Finetuning Models' guide for adapting pretrained weights.

Common use cases and workflows

• •Training a custom object detector on your own COCO-format dataset for a specific class list.
• •Swapping in a different backbone (Swin Transformer or ConvNeXt, for example) on an existing detection head to compare accuracy against speed without rewriting the training loop.
• •Running instance segmentation with Mask R-CNN or newer methods like Mask2Former from the same codebase used for detection.
• •Real-time or rotated object detection with RTMDet, per its reported benchmarks on COCO and DOTA.
• •Open-vocabulary or phrase-grounding detection via MM-Grounding-DINO.

MMDetection vs Detectron2, MMYOLO, and other alternatives

Frequently asked questions

The problem it solves

Object detection research produces a new architecture every few months — a new backbone, a new detection head, a new loss — and reimplementing each one from its paper to compare against what you're already running is its own project. MMDetection's problem is that comparison gap: it reimplements a long list of published detectors (from Fast R-CNN to RTMDet to DINO) inside one shared codebase so switching between them is a config change rather than a from-scratch reimplementation.

Who should try it — and who should skip

Try MMDetection if you're a researcher or engineer who needs more than one detection architecture over a project's lifetime — comparing Faster R-CNN against RTMDet against a transformer-based DETR variant is a config swap here, not a rewrite. It also suits teams already inside the OpenMMLab ecosystem (MMCV, MMEngine) or building on MM-Grounding-DINO for open-vocabulary detection. Skip it if you want one YOLO model running in an afternoon with minimal dependency overhead — the config system and MMEngine/MMCV stack are more machinery than a single-architecture project needs.

Related repositories