mmdetection
**MMDetection** is an **open-source object detection toolbox built on PyTorch that provides a comprehensive model zoo of hundreds of detection algorithms with a modular, configurable architecture** — part of the OpenMMLab project, it decomposes detection frameworks into interchangeable components (backbone, neck, head, RoI extractor) that researchers can mix and match to create new architectures, reproduce published results, and benchmark detection methods on a level playing field.
**What Is MMDetection?**
- **Definition**: A PyTorch-based detection framework from the OpenMMLab ecosystem (Chinese University of Hong Kong) that implements 300+ detection models and 50+ datasets in a unified codebase — providing config-driven training where switching from Faster R-CNN to DETR requires changing a config file, not rewriting code.
- **Model Zoo**: The most comprehensive collection of detection algorithm implementations — Faster R-CNN, Mask R-CNN, Cascade R-CNN, YOLO series, SSD, RetinaNet, FCOS, ATSS, DETR, Deformable DETR, DINO, Co-DETR, and dozens more, all with pretrained weights and benchmark results.
- **Modular Design**: Detection models are decomposed into standardized components — Backbone (ResNet, Swin Transformer, ConvNeXt), Neck (FPN, PAFPN, BiFPN), Dense Head (anchor-based, anchor-free), RoI Head (RoI Align, RoI Pool) — each swappable via config.
- **Config System**: Models are defined entirely in Python config files — inherit from base configs, override specific components, and compose complex architectures without touching source code.
- **Research Standard**: The default framework for publishing detection papers — researchers implement their method in MMDetection to ensure fair comparison with existing methods on standard benchmarks.
**Key Features**
- **300+ Models**: Every major detection architecture from 2015-2025 — two-stage (Faster R-CNN family), single-stage (YOLO, SSD, RetinaNet), anchor-free (FCOS, CenterNet), and transformer-based (DETR, DINO).
- **Benchmark Reproducibility**: Every model config includes expected mAP on COCO — researchers can verify their setup reproduces published numbers before modifying the architecture.
- **Training Recipes**: Optimized training schedules (1x, 2x, 3x) with learning rate warmup, multi-scale training, and test-time augmentation — following community best practices.
- **Distributed Training**: Native support for multi-GPU and multi-node training via PyTorch DDP — scale training to large datasets and complex models.
- **Inference Pipeline**: `MMDetInferencer` provides a simple API for loading any model and running inference on images, videos, or webcam streams.
**MMDetection Architecture Components**
| Component | Role | Examples |
|-----------|------|---------|
| Backbone | Feature extraction | ResNet-50, Swin-T, ConvNeXt-B |
| Neck | Feature fusion | FPN, PAFPN, BiFPN |
| Dense Head | Proposal/detection | RPN, RetinaHead, FCOSHead |
| RoI Head | Region refinement | StandardRoIHead, CascadeRoIHead |
| Loss | Training objective | CrossEntropy, FocalLoss, GIoU |
| Data Pipeline | Augmentation | Mosaic, MixUp, RandomFlip, Resize |
**MMDetection vs Alternatives**
| Feature | MMDetection | Detectron2 | Ultralytics YOLO | torchvision |
|---------|-----------|-----------|-----------------|-------------|
| Model count | 300+ | 50+ | YOLO family only | 10+ |
| Research focus | Excellent | Excellent | Production | Basic |
| Config system | Python configs | YAML (Detectron2) | YAML | Code-only |
| Ease of use | Moderate | Moderate | Excellent | Easy |
| Community | Very active (OpenMMLab) | Active (Meta) | Very active | PyTorch core |
| Paper reproduction | Standard | Common | Rare | Rare |
**MMDetection is the research-grade detection toolbox that provides the most comprehensive collection of detection algorithms in a single unified framework** — enabling researchers to fairly benchmark new methods against hundreds of existing approaches and practitioners to quickly prototype detection systems using battle-tested implementations of every major architecture.