[Paper Review] MMDetection: Open MMLab Detection Toolbox and Benchmark

Kai Chen, Jiaqi Wang+23 more|arXiv (Cornell University)|Jun 17, 2019

Advanced Neural Network Applications41 references794 citations

TL;DR

MMDetection is a PyTorch-based, modular object detection and instance segmentation toolbox with a large model zoo and a comprehensive benchmark, including ablations and cross-codebase comparisons."

ABSTRACT

We present MMDetection, an object detection toolbox that contains a rich set of object detection and instance segmentation methods as well as related components and modules. The toolbox started from a codebase of MMDet team who won the detection track of COCO Challenge 2018. It gradually evolves into a unified platform that covers many popular detection methods and contemporary modules. It not only includes training and inference codes, but also provides weights for more than 200 network models. We believe this toolbox is by far the most complete detection toolbox. In this paper, we introduce the various features of this toolbox. In addition, we also conduct a benchmarking study on different methods, components, and their hyper-parameters. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible toolkit to reimplement existing methods and develop their own new detectors. Code and models are available at https://github.com/open-mmlab/mmdetection. The project is under active development and we will keep this document updated.

Motivation & Objective

• Provide a unified, modular framework for constructing and training object detectors.
• Benchmark a wide range of detection methods and components to enable fair comparisons.
• Offer practical training, ablation, and hyper-parameter insights to guide researchers.
• Present a lightweight, efficient, and scalable toolbox with extended model weights and multi-framework support.

Proposed method

• Introduce a modular detector architecture with Backbone, Neck, DenseHead, RoIExtractor, and RoIHead abstractions to enable easy composition of detectors.
• Provide a unified PyTorch-based training pipeline with a hook system for flexible customization.
• Benchmark multiple detection frameworks and backbones on COCO 2017 to compare accuracy, speed, and memory.
• Conduct extensive ablation studies on losses, normalization layers, training scales, and other hyper-parameters.
• Compare MMDetection against Detectron, maskrcnn-benchmark, and SimpleDet in terms of speed, memory, and AP on shared hardware.

Experimental results

Research questions

• RQ1 How does a modular, reusable component design affect detector development and reproducibility?
• RQ2 What are the performance and efficiency trade-offs across modern detectors and backbones on COCO?
• RQ3 How do different losses, normalization strategies, and training scales influence object detection performance?
• RQ4 How does MMDetection compare to other codebases in speed, memory, and accuracy under equivalent conditions?

Key findings

• MMDetection provides a broad set of models and over 200 network weights, enabling flexible reimplementation and development.
• Mixing precision training (FP16) reduces memory with minimal accuracy loss and yields speed benefits on modern GPUs.
• Normalization strategies (FrozenBN, SyncBN, GN) and head design choices significantly impact AP, with deeper bbox heads and alternative head architectures improving performance.
• Training scale and data augmentation (range vs value multi-scale) modestly improve AP, with wider ranges providing gains at larger scales.
• IoU-based and alternative regression losses show varying gains depending on loss weights and sampling strategies, with some losses outperforming Smooth L1 under tuned conditions.
• MMDetection achieves near-linear speedups with multi-node training and demonstrates competitive performance across multiple backbones and detectors.