CS231n Lecture 11 강의 정리

Semantic Segmentation Idea: Fully Convolutional

• Downsampling: Pooling, strided, convolution
• Upsampling: ???

In-Network upsampling: "Unpooling"

In-Network upsampling: "Max Unpooling"

Leanable Upsampling: Transpose Convolution

Transpose Convolution: 1D Example

Semantic Segmentation Idea: Fully Convolutional

• Downsampling: Pooling, strided, convolution
• Upsampling: Unpooling or strided transpose convolution

Classification + Localization

Multitask Loss

Often pretrained on ImageNet (Transfer learning)