1. Fully Convolutional Networks for Semantic Segmentation

1. VGG backbone
2. FC layer → Convolution
3. Transposed Convolution

Transposed Convolution

• Upsampling = Deconvolution = Transposed Convolution
• Transposed Convolution parameters are learnable parameters
• The precise term is “Transposed Convolution”, not “Deconvolution” as this operation does not output the same matrix as the input of the convolution operation

FCN

1. MaxPooling에 의해서 잃어버린 정보를 복원해주는 작업 진행
2. upsampled Size를 줄여주기에 좀 더 효율적인 이미지 복원이 가능

Evaluation Metric

Pixel Accuracy