参考论文：《U-net: Convolutional networks for biomedical image segmentation》 lmb.informatik.uni-freiburg.de/people/ronn…
| max pooling | stride=2 |
| others | stride=1 (unpaddied) |
- based on Fully Convolutional Network
- contracting path; expansive path;
- The desired output include localization
- The upsampling part we have also a large number of feature channels -> higher resolution
- CHALLENGE: The separation of touching objects of the same class.
- upsample and downsample
- At each downsampling setp we double the number of feature channels
- At each upsampling step we halves the number of feature channels
- each 3x3 convolution followed by a ReLU
- SGD
- high momentum: 0.99;
最后输出，有K个channel，每个Channel代表一个类别。
对于ground truth，相同类别存在相连的情况：
(在处理同类别连接边界问题和groud truth权重计算方面，有些不理解）
