Home

Introduction

MMDetection is a toolkit for object detection. It consists of the following modules.

• core: core functionality such as anchor generation, training target computing, etc.
• datasets: coco-style dataset, voc-style dataset, self-defined pickle dataset
• models: components and detectors
• ops: custom operators, mostly cuda extensions.
• tools: training/testing/evaluation/conversion.

Release Plan

v0.5.0 (before 07/09)

The first release will contain implementations of RPN, Faster R-CNN, Mask R-CNN, FPN, Cascade R-CNN. Only COCO dataset is supported.

• refactoring and finalize the API (01/09)
• datasets.coco (02/09)
• models.backbones.bricks, models.backbones.resnet (02/09)
• models.necks.fpn (03/09)
• models.rpn_heads (03/09)
• models.bbox_heads (03/09)
• models.mask_heads (03/09)
• ops.roi_pooling, ops.roi_align (03/09)
• models.detectors.rpn, models.detectors.two_stage, models.detectors.cascade (04/09)
• tools.train, tools.test (05/09)
• documentation (06/09)

v0.6.0 (before 20/09)

• Add ResNeXt backbone.
• Add support for Fast R-CNN
• Add support for single-stage detector
• Add VOC-style dataset and self-defined pickle-based dataset

v0.7.0 (before 30/09)

• Add deformable convolution
• Add SyncBN and GN
• Add support for R-CNN
• Add support for R-FCN