imaginea si forma 3D nu se suprapun

• aggregate model summaries (HG, U-Net, VGG, VRN-Unguided, FAN)
• visualize model output in meshlab via octave
  • why is it so sparse and long on z?
• use a feature extractor trained on faces
  • also try with their FAN

• landmark heatmaps
  • check if landmarks from dataset are aligned with the volume
• double check u-net <-> hourglass arch
  • implement standalone hourglass module
• model output visualization
  • construct denser representation
  • marching-cubes
• overfit smaller dataset
• write dataloader (3D facial scan -> 3D binary volume)
• generate 3D coords from .mat (where not present)
• convert torch model to pytorch and visualize it
• data augmentation (transformations on labels)

Tried

• smaller u-net + stacked u-net
• bigger 2xHourglass
• predict depth-map instead of volume
  • feature extraction (VGG) + 2xFC (relu + dropout)
  • additional loss term
• closing + gaussian smoothing
• vrn_unguided from torch (with UpsamplingBilinear2d / ConvTranspose2d)
• weighted loss on gaussian landmarks