Un-Supervised

• Auto-encoder FCN
• Learns bottleneck (latent) representation
  • Information rich
  • $f(.)$ is CNN encoding function

Supervised

• Triplet loss
  • Providing positive and negative requires supervision
• Two losses