mmdetection

MMDetection is an open-source object detection toolbox built on top of PyTorch. It is developed by OpenMMLab and provides a comprehensive, modular framework for training and evaluating object detection models. The project is designed so that different parts of a detection pipeline — the backbone network, the neck (feature pyramid), the head (which makes predictions), and the loss functions — are all independently configurable, making it easy to mix and match components or plug in new ones.

The toolbox ships with implementations of a large number of detection algorithms, including both two-stage detectors like Faster R-CNN and Cascade R-CNN, and single-stage detectors like FCOS and RetinaNet, as well as newer transformer-based approaches like DETR and its variants. Many common benchmarks and their pretrained weights are provided, allowing researchers to start from a strong baseline rather than training from scratch.

MMDetection handles the full training pipeline: data loading with augmentation, distributed multi-GPU training, logging, checkpointing, and evaluation against standard metrics like COCO mAP. Configuration is managed through Python config files that fully specify the model, dataset, training schedule, and evaluation settings.

When to use it: MMDetection is aimed at computer vision researchers and engineers who need to train or fine-tune object detection models. It is most useful when experimenting with detection architectures, comparing algorithms on a consistent codebase, or adapting existing models to a custom dataset. It is less suited for production inference deployments where a lighter runtime like ONNX or TensorRT integration would be preferred.

The tech stack is Python and PyTorch throughout.