부스트캠프 14주차 학습 일지 - Semantic Segmentation

사용한 기술 스택들:    

12/19 월

학습한 것들:

mIOU: mean of IOU over classes

 

FCN: use the VGG network as the backbone and replace the FC layer with Convolution

• allows using pretrained networks for better performance
• pixel-wise prediction
• convolution is irrelevant to image size
• transposed convolution for upsampling
• skip connection to have a sharp image
• Small objects are often ignored

DeconvNet: make the encoder and decoder symmetrical by using unpooling and deconvolution

• Unpooling: save the edge that was deleted during pooling
  
  • fast as it does not need to be trained
  • has a sparse activation map so it is required to use transposed convolution too
  • unpooling captures "example-specific" structure
  • transposed convolution captures a "class-specific" structure

SegNet: use convolution to turn the sparse matrix into a dense matrix

 

FC DenseNet: add skip connection inside a block and from the decoder to the encoder

 

DeepLab v1: mixing convolution and max pooling can increase the receptive field of a unit pixel of the feature map but is low in quality

• use dilated convolution to be more efficient
• use bilinear interpolation for upsampling
• dense conditional random field to get pixel-wise segmentation

DilatedNet: DeepLab but only 2x2 max pool in the beginning

 

---

12/21 수

학습한 것들:

DeepLab v2: added ASPP which are branches that replaced the fully convolutional layer

• branches act as an ensemble
• ResNet backbone

PSPNet: use global average pooling to take surroundings into account

• Accounts for mismatched relationship
  • takes the surrounding into account
• Confusion categories
  • similar category can be confusing
• Inconspicuous classes
  • small objects can be detected using global contextual information

DeepLab  v3: added global average pooling with 1x1 convolutional network

 

DeepLab v3+: use encoder-decoder structure again

• use a decoder to restore the lost spatial information
• modified Xception Backbone
  • use depthwise + pointwise convolution