choose architectures

[pesekon2]

Research on what CNN architectures test should be done.

Proposed architectures and in what paper they were used:

• [x] U-Net (this review)
• [x] Segnet (this review)
• [ ] FC-DenseNet (this review)
• [x] FCN (probably shoulnd't miss this one, according to this review, it is the only one we're missing from the big four)
• [x] DeepLabV3+ (this review, Michal's paper)
  • [x] Backbone: ResNet-50
  • [x] Backbone: ResNet-101
  • [x] Backbone: ResNet-151
  • [ ] Backbone: Xception with atrous separable convolutions
• [ ] PSPNet (this review)
• [ ] MobileU-Net (this review)
• [ ] CloudFCN (this paper)
• [ ] MCNN (this paper)
• [ ] FECN (this paper - seems slightly wicked to me)
• [ ] P_SegNet and NP_SegNet (this paper) - sort of a bullshit one with unsatisfying reults. Skip if not needed
• [ ] Bi-SAD (this paper)
• [ ] MF-CNN (this paper)
• [ ] MSCFF (this paper) - looks actually promising
• [ ] MFFSNet
• [ ] CDNet
• [ ] HRNet (this paper)
• [ ] ACABFNet (this paper)
• [ ] FCNet (FC-EF and some Siam thingeys) (this paper)
• [ ] SuaCDNet (for change detection) ((this paper)[https://www.sciencedirect.com/science/article/pii/S0303243421003044?via%3Dihub])