DeepLabv3+训练细节总结

Training Protocol

• backbone: ResNet-101 or modified aligned Xception
• pretrain: ImageNet-1K
• dataset: PASCAL VOC 2012 (20 foreground object classes, 1 background class)
  10582 (trainaug) training images, 1449 (val), 1456 (test)
• lr schedule: “poly” policy (initial lr: 0.007)
  initial learning rate is multiplied by $(1-\frac{iter}{max\_iter}{)}^{power}$ with $power=0.9$
• crop size: 513×513
  For atrous convolution with large rates to be effective, large crop size is required.
• fine-tune batch normalization parameters when output stride= 16
  output stride: the ratio of input image spatial resolution to final output resolution.
  added modules (ASPP, decoder, etc) on top of ResNet all include batch normalization parameters.
• batch size=16
  The batch normalization parameters are trained with decay = 0.9997.
  After training on the trainaug set with 30K iterations and initial learning rate = 0.007, we then freeze batch normalization parameters, employ output stride = 8, and train on the official PASCAL VOC 2012 trainval set for another 30K iterations and smaller base learning rate = 0.001.
• random scale data augmentation: scaling input image (from 0.5 to 2.0) and randomly left-right flipping during training
• include batch normalization parameters in the proposed decoder module
• train end-to-end
• Upsampling logits

Inference strategy on val set

• output stride=8
  the model is trained with output stride=16, and apply output stride=8 to get more detailed feature map during inference.