model | mAP | eval method | Training set | PreTrain set | Training Log | Eval Log | Base Module | config |
---|---|---|---|---|---|---|---|---|
resnet101+deepmask | 0.5090 | nms+box voting | Imagenet all | - | - | - | - | NMS=0.5 IoU_Thresh=0.5 score_Thresh=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 |
- 7 * BN-Inception (32 Layers)
- 2 * MSRA-Net (22 Layers)
- ResNet, Identity Map
- random crop
- multi-scale
- contrast jittering
- color jittering
- Pretrain on LOC !!
- Objectness loss
- Negative categories
- BBox Voting
- Balanced Sampling
- Multi-Scale Training
- Online Hard Sample Mining
- Cascade RPN
- Constrained Neg/Pos Anchor Ratio
- Pretrained Global Context
- 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