[CV] Semantic segmentation

Sementic segmentation

- A classification peoblem for each pixel

- Not about classifications of instances.

 

Fully Convolutional Networks(FCN)[Long et al., CVPR 2015]

- the first end-to-end architecture for semantic segmentation

- Convolution networks without a flatten layer(input resolution free)

- 1x1 convolution : the same operation with fc layer for each spatial point.

- low resolution of feature map : large receptive -> upsample and apply

- upsampling

  - transposed convolution has overlap issues.

  - Intrgrates activations from lower layers into prediction

    - near input : Fine Low-level Detail Local

    - near output : Coarse Semantic Holistic Global

 

Hypercolumns for object segmentation[Hariharan et al., CVPR 2015]

- integrates activation from lower layers into perdiction

 

U-Net[Ronneberger et al., MICCAI 2015]

- Contracting path and Expanding path

- Contracting path

  - convolution layer and ReLU activation

  - maxpooling(compressing spatial dimension and expanding channel dimension)

-Expanding path

  - convolution transposed layer(compressing channel dimension and expanding spatial dimension)

  - kernel size = 2, stride = 2 : to avoid checker board artifact

 

DeepLab[Chen et al., ICLR 2015]

- CRF : post-processes a segmentation map to be refined to follow image boundaries.

- Dilated convolution

  - large receptive field

  - Depthwise separable convolution (proposed by Howard et al.)

- DeepLab v3+[Chen et al., ECCV 2018]

  - Atrous spatial pyramid pooling

  - Decoder module and upsample

 

###

 

피어세션

 

https://www.notion.so/vgg11-Batch_norm-2280d43fa25e45aaa40c15c035e1f893

vgg11의 Batch_norm 유무