zenozhouzhao / detectiontricks

ImageNet Detection Tricks

ImageNet Detection Tricks

ATLAB ImageNet leaderboard

model	mAP	eval method	Training set	PreTrain set	Training Log	Eval Log	Base Module	config
Resnet-101 std	0.4874	test.py	ImageNet all	None	Train	Test	Resnet-101 param Resnet 101 modeljson	config
ResNet-101 smalldb	0.3958	test.py	ImageNet train_0	None	Train	Eval	Resnet-101 param Resnet 101 modeljson	config

基础模型

• ResNet-101 Variants

使用多个差异很大的CNN模型 - diversity matters!

• 7 * BN-Inception (32 Layers)
• 2 * MSRA-Net (22 Layers)
• ResNet, Identity Map

数据放大

• random crop
• multi-scale
• contrast jittering
• color jittering
• Pretrain on LOC !!

单个模型的改进

• Objectness loss
• Negative categories
• BBox Voting

训练技巧

• Balanced Sampling
• Multi-Scale Training
• Online Hard Sample Mining

RPN Proposal

• Cascade RPN
• Constrained Neg/Pos Anchor Ratio

Pretraining

• Pretrained Global Context

测试技巧

• Multi-Scale Testing
• HFlip
• Box Votinng

Tricks的实现划分在以下5个文件夹中：

• dataprocess
• regionproposal
• fastrcnn
• postprocess
• ensumble

上述代码尝试做成平台无关，与计算框架相关的代码都在PlatformRelated文件夹中