Overview: This paper summarized the existing deep learning approaches in change detection and apply UNet++ to sattlie image with MSOF(Multiple Side-Outputs Fusion).
UNet++: On top of UNet, UNet++ re-designed skip connections. Firstly, it has dense connection on the skip pathways (a comparison of DenseNet v.s. ResNet), which might improve the gradient flow. Also, along the skip pathways, it has convolution layers to bridge the semantic gap between encoder and decoder feature maps.
Deep-supervision: Instead of the supervise the mode from the very end (loss function), we can supervised at differnt semantic levels; and UNet++ provides mutiple side outputs. This paper implement the deep-supervision a bit different from the original one: It concatenate 4 side outputs and sum up the loss for all outputs
Loss Function: The loss function is a combination of balanced binary cross-entropy loss and dice loss.
Result: No experiments on established dataset (such as COCO) are carried out.
Q1: Can you draw a simple diagrams of FB-DLCD, PB-DLCD and IB-DLCD? Any comments on the system design?
Follow up: Why the author said the traditional CD methods works on the assumption the number of changed pixels is propotional to that of the unchanged?
Q2: How is the UNet++ architecture different from UNet? Could you compare figure 1a) with the UNet diagram? What are some advantage of this design?
Q3: What is Deep Supervision (DS) and Multiple Side-Outputs Fusion (MSOF)? How are they implemented in UNet++?
Q4: What is the lambda doing between BCE and Dice loss?
Overall comments: Not a lot of novelty, but balanced and used the right tools to address the application problem.
Note most of the architecture setting are proposed in the orignal UNet++ paper