ImageNet Detection Tricks

ATLAB ImageNet leaderboard

| model | mAP | eval method | Training set | PreTrain set | Training Log | Eval Log | Base Module| config | | ----- |---|---|---|---|---|---|---|---| | 【ALL】= resnet101 (multiscale) + resnet101 ratio1:4,4:1 (multiscale) + resnet101 scale4 (multiscale) + resnet152 (multiscale) + inceptionv3 (multiscale) + rcnndcn (multiscale) | 0.5295 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS1】=【ALL】 - dcn_rcnn (scale1000, mAP=0.4231) | 0.5301 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS2】=【ENS1】 - dcn_rcnn (scale400, mAP=0.4249) | 0.5297 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS3】=【ENS2】 - inceptionv3 (scale400, mAP=0.4257)| 0.5289 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS4】=【ENS3】 - resnet152 (scale400, mAP=0.4318) - resnet101_ratio4 (scale400, mAP=0.4350) - resnet101_scale4 (scale400, mAP=0.4375)| 0.5289 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS5】=【ALL】 - resnet152 (scale400, mAP=0.4318) - resnet101_ratio4 (scale400, mAP=0.4350) - resnet101_scale4 (scale400, mAP=0.4375)| 0.5298 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS6】=【ENS1】 + dcn_rfcn (scale600, mAP=0.4695) | 0.5305 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS7】=【ENS6】 - resnet101 (scale400, mAP=0.4456) | 0.5295 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS8】=【ENS7】 - resnet101 (scale1000, mAP=0.4532) | 0.5301 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS9】=【ENS8】 - resnet101_scale4 (scale1000, mAP=0.4555) | 0.5296 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | 【ENS10】=【ENS9】 - dcn_rfcn (scale800, mAP=0.4597) | 0.5292 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | resnet101(multiscale) + resnet152 + inceptionv3 + rcnndcn | 0.5238 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | resnet101(multiscale) +resnet152 +inceptionv3 | 0.5222 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | resnet101(multiscale) +deepmask +resnet152 +inceptionv3 | 0.5213 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | resnet101(multiscale) +deepmask +resnet152|0.5141 |nms+box voting | Imagenet all|-|-|-|-|NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | | resnet101(multiscale) +deepmask | 0.5090 | nms+box voting | Imagenet all | - | - | - | -| NMS＝0.5 IoUThresh=0.5 scoreThresh=0.1 | |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