Unsupervised Learning of Object Landmarks through Conditional Image Generation

learn landmark by looking objects from diff. vp

• a mapping gamma(x, phi(x’)) = x’
• phi(x’) being a set of landmark detections
  • avoid identical mapping
• heatmap -> renormalised to prob. dist.
  • softmax then condensed to a point
• perceptual loss
  • pretrained by colorising grayscale images