单阶段目标检测器梳理
- 1、Training
- 1.1、Data augmentation
- 1.1.1、Multi-scale
- 1.1.2、Random crop
- 1.1.3、Color jitter
- 1.1.4、Mixup
- 1.1.5、Label smoothing
- 1.1.6、Random erase
- 1.2、Model
- 1.2.1、Backbone
- 1.2.1.1、ResNet
- 1.2.1.2、MobileNet series
- 1.2.2、Neck
- 1.2.2.1、FPN
- 1.2.2.2、PAFPN
- 1.2.2.3、NASFPN
- 1.2.2.4、BiFPN
- 1.2.3、Head
- 1.2.3.1、Shared between feature map lavel or not
- 1.2.4、Misc
- 1.2.4.1、Normalization
- 1.2.4.1.1、Batch norm
- 1.2.4.1.2、Group norm
- 1.2.4.2、Weight initialization
- 1.2.4.2.1、Uniform、xavier、gaussian、kaiming
- 1.2.4.2.2、Special initialization like the way in focal loss
- 1.2.4.3、Gradient clipping
- 1.2.1.1、ResNet
- 1.3、Positive/negative/ignore labels assignment
- 1.3.1、Feature map level selection
- 1.3.2、Point/bbox label assignment
- 1.4、Loss
- 1.4.1、Classification
- 1.4.1.1、Cross entropy loss
- 1.4.1.2、Binary cross entropy loss
- 1.4.2、BBox Regression
- 1.4.2.1、Distance based
- 1.4.2.1.1、L2 loss
- 1.4.2.1.2、SmoothL1 loss
- 1.4.2.1.3、L1 loss
- 1.4.2.2、IoU based
- 1.4.2.2.1、GIoU loss
- 1.4.2.2.2、DIoU loss
- 1.4.2.2.3、CIoU loss
- 1.4.3、Imbalance
- 1.4.3.1、OHEM
- 1.4.3.2、GHM
- 1.4.3.3、Weighted loss
- 1.4.3.4、Focal loss
- 1.4.4、Misc
- 1.4.4.1、Normalization
- 1.4.4.1.1、Sample-wise
- 1.4.4.1.2、Batch-wise
- 1.4.4.1.3、None
- 1.4.4.1、Normalization
- 1.4.1、Classification
- 1.5、Optimizer
- 1.5.1、SGD
- 1.5.2、Adam
- 1.5.3、RMSprop
- 1.6、LR scheduler
- 1.6.1、MultiStepLR
- 1.6.2、CosineAnnealingLR
- 1.6.3、CyclicLR
- 1.6.4、Misc
- 1.6.4.1、Warmup
- 1.2.1、Backbone
- 1.1、Data augmentation
- 2、Testing
- 2.1、Test time augmention
- 2.1.1、Multi-scale
- 2.1.2、Flipping
- 2.2、Model ensemble
- 2.2.1、Bagging
- 2.2.1.1、Averaging
- 2.2.1.2、Majority voting
- 2.2.1、Bagging
- 2.3、Postprocessing
- 2.3.1、NMS、soft-NMS
- 2.3.1.1、Class aware
- 2.3.1.2、Class agnostic
- 2.3.1、NMS、soft-NMS
- 2.4、Confidence scores interpreting
- 2.4.1、Process each class separately or not
- 2.1、Test time augmention
欢迎来GitHub Discussions讨论