Original article can be found here (source): Deep Learning on Medium
Edge Guided GANs with Semantic Preserving for Semantic Image Synthesis
The semantic mask → creates realistic images → these are generative models → and super interesting work → how much of details can the model add?
They are going to use edges as a signal to make the model more effective → when it comes to image generation.
Quite a lot of networks are used → to actually make this happen → the complex yet very effective way of creating new images → interesting…
The problem really seems to be → can GAN create all of the objects → not in the semantic mask → how would it know what kind of object to create.
They were the first when it comes to creating → the images while using edges as leverage → a good idea. (they are bringing innovation to the field).
Conditional GAN → combines the label → generate while taking certain conditions in considerations. (a good idea → create what we want from a condition).
Different kinds of modules are used → to better the process of learning a manifold of data → and we can map more items → the key idea is to make an efficient function for both encoding and decoding.
There are many more → than just GAN → five modules that are used to solve this problem → , not an easy thing to do for smaller computers. (Canny Edge Detector is used → to create ground truth edges).
A lot of feature binding and → concatenation → a system where past regressions can affect the final results.
LOOL → wow very complex shit LOL → at final step → sigmoid is applied to normalize the value between ranges.
Again we can see that loss function is tuned → to get the model working in the way we want. (and they really used famous segmentation dataset → real-world dataset → and got impressive results!).
There are a little bit of artifact → and this is understandable → it is really hard to perfect the small details → not an easy task to do.
Quite surprised on the light changes → they model is actually able to put in light changes of the floor. (and edges really does help → for better generation).
Important research was done → edges are good for a generation.