Automatically generate and complete the streetview

• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• Decode the simple(uniform here) distribution p_z to the images(streetview here) distribution p_data
• 
• How to use: DCGAN_streetview/ main.py --dataset datasetname --mode train
• TODO:
  • GAN-improved openai/improved-gan
  • VAE
  • step-by-step GAN

Now we can pick an arbitrary z~p_z and decode it to a image

• Semantic Image Inpainting with Perceptual and Contextual Losses
• Pick the z that fits the original image well
• 
• Minimizing the (Contextual + lamda*Perceptual) loss
• How to use: DCGAN_streetview/ main.py --dataset datasetname --mode complete
• TODO:
  • Specific generative model?

• Poisson Image Editing
• Synthesize new image with poisson blending
• 
• How to use: poissonblending/ main.py
• TODO:
  • Heat map

• related to Generative Image Modeling using Style and Structure Adversarial Networks
• Here is another example which focus on pedestrian.
• In the early training stage, the network seems to decide the structures or poses of pedestrians. Then in the late training stage, it only has subtle changes on texture according to the current batch.
• 
• In the completion stage, it tends to choose those z resulting blurry images
• 
• TODO:
  • Inverse mapping to the latent space for GAN

• If we can complete the missing part of a image, can we combine one generator(pedestrian) with another discriminator(streetview)?

• Cityscapes labels