商汤科技和香港中文大学开源的基于Pytorch实现的深度学习目标检测工具箱。

MMDetection的特点：

1. Architecture

目标检测模型通常由不同的组件构成。MMDetection中的基本组件包括：

Backbone：把image转换成feature map，如ResNet-50；
Neck：连接backbone和head，对raw feature map进行refine和reconfigure，如FPN；
DenseHead (AnchorHead/AnchorFreeHead)：进行dense location，包括AnchorHead和AnchorFreeHead，如RPN-Head、RetinaHead、FCOSHead;
RoIExtractor：提取RoI(Region of Interests)，如RoIPooling、RoIAlign；
RoIHead (BBoxHead/MaskHead)：使用RoI进行预测，如bbox分类/回归、mask预测。

用基本组件组成的Single-stage和Two-stage的模型框架：

训练（training epochs）和验证（validation epochs）时设计了hooking mechanism。训练时的pipeline如下，验证类似。

2. Hyper-parameters

目标检测的损失包括边界框回归损失和分类损失。可选的回归损失如下：

不同的回归损失和损失系数搜索如下：

目标检测训练时batch较小，使用Batch Normalization时设置如下：

可选的Normalization：

不同的Normalization及其应用的位置搜索如下：

目标检测默认的训练scale为1333×800。

MMDetection允许multi-scale训练，即每次训练时选择一个输入image的scale，提供了两种scale选择方法：

不同的training scale搜索如下：

其他超参数包括：

smoothl1 beta：Smooth L1 Loss的参数torch.where(x < beta, 0.5·x^2/beta, x - 0.5·beta)
allowed border：边界框超出原图像的允许上界
neg pos ub：negative和positive样本比例

这些超参数搜索如下：

SSD： a classic and widely used single-stage detector with simple model architecture
RetinaNet： a high-performance single-stage detector with Focal Loss
GHM: a gradient harmonizing mechanism to improve single-stage detectors
FCOS: a fully convolutional anchor-free singlestage detector
FSAF: a feature selective anchor-free module for single-stage detectors

Fast R-CNN: a classic object detector which requires pre-computed proposals
Faster R-CNN: a classic and widely used twostage object detector which can be trained end-to-end
R-FCN: a fully convolutional object detector with faster speed than Faster R-CNN
Mask R-CNN: a classic and widely used object detection and instance segmentation method
Grid R-CNN: a grid guided localization mechanism as an alternative to bounding box regression
Mask Scoring R-CNN: an improvement over Mask R-CNN by predicting the mask IoU
Double-Head R-CNN: different heads for classi-fication and localization

Cascade R-CNN: a powerful multi-stage object detection method
Hybrid Task Cascade: a multi-stage multi-branch object detection and instance segmentation method

Mixed Precision Training: train deep neural networks using half precision floating point (FP16) numbers
Soft NMS: an alternative to NMS
OHEM: an online sampling method that mines hard samples for training
DCN: deformable convolution and deformable RoI pooling
DCNv2: modulated deformable operators
Train from Scratch: training from random initialization instead of ImageNet pretraining
ScratchDet: another exploration on training from scratch
M2Det: a new feature pyramid network to construct more effective feature pyramids
GCNet: global context block that can efficiently model the global context
Generalized Attention: a generalized attention formulation
SyncBN: synchronized batch normalization across GPUs
Group Normalization: a simple alternative to BN
Weight Standardization: standardizing the weights in the convolutional layers for micro-batch training
HRNet: a new backbone with a focus on learning reliable high-resolution representations
Guided Anchoring: a new anchoring scheme that predicts sparse and arbitrary-shaped anchors
Libra R-CNN: a new framework towards balanced learning for object detection

一些frameworks的表现如下：