[AIFFEL] 22.Apr.08, GD_OCR_3(Segmenatation만 한 날...)

논문 구현

U-Net ++

• Redesigned skip connections

  • skip connections와 upsampled lower-level feature map을 concatenate 할 때 채널 수가 어떻게 되는지가 관건인데

  • p 2. "As a result, each node in the UNet++ decoders, from a horizontal perspective, combines multiscale features from its all preceding nodes at the same resolution, and from a vertical perspective, integrates multiscale features across different resolutions from its preceding node, as formulated at Eq. 1. This multiscale feature aggregation of UNet++ gradually synthesizes the segmentation"

  • p 2. "We redesign skip connections in UNet++, enabling flexible feature fusion in decoders—an improvement over the restrictive skip connections in U-Net that require fusion of only same-scale feature maps (see Section II-B)."

  • p 3. "While using aggregated feature maps at a decoder node is far less restrictive than having only the same-scale feature map from the encoder, there is still room for improvement. We further propose to use dense connectivity in UNet+, r"

  • p 3. "With dense connectivity, each node in a decoder is presented with not only the final aggregated feature maps but also with the intermediate aggregated feature maps and the original same-scale feature maps from the encoder. As such, the aggregation layer in the decoder node may learn to use only the same-scale encoder feature maps or use all collected feature maps available at the gate."

  • p 3. "Let $x^{i,j}$ denote the output of node $X^{i,j}$ where $i$ indexes the down-sampling layer along the encoder and $j$ indexes the convolution layer of the dense block along the skip connection."

• X_00, X_01, X_02, X_03, X_04 끝처리 부분

  • p 4. "For this purpose, we append a 1 x 1 convolution with $C$ kernels followed by a Sigmoid activation function to the outputs from nodes $X^{0,1}$, $X^{0,2}$, $X^{0,3}$, and $X^{0,4}$ where $C$ is the number of classes observed in the given dataset."

• Loss

  • 

• Deep Supervision

  • 이거 근데 뭘까...? 예측으로는 뭔가 branch들 하나하나 Supervision(훈련?) 해줬어야 했는데, 지금은 안해도 된다는 건가...?
  • p 2. "Note that, explicit deep supervision is required (bold links) to train U-Nete but optional (pale links) for UNet+ and UNet++."