ToDo

[vikasTmz]

• [x] GCP : group project + get access to better GPUs
• [x] Visualize TextureFields results: Find out what's good and bad about their results.
• [x] Write up and submit Check in 2
• [x] Email Eliot to set up next meeting/start email chain
• [ ] Understand TextureFields code (which files contain what will be a helpful start).
• [ ] Setup DeepSDF : get pre-trained ShapeNet weights / train it on 3D Future.
• [ ] Preprocess 3D Future to match ShapeNet's data format.

[vikasTmz]

Possible paths:

• [ ] Use a pre-trained PointNet encoder (on cars) from the occupancy network paper instead of training the PointNet encoder from scratch.

  • Experiments will be done on ShapeNet cars.

• [ ] Replace/ append another loss instead of just l1 (L_conditional).

  • Experiments can be done on either dataset.
  • Try using a perceptual loss maybe?

• [ ] Condition the discriminator in the cGAN setup with more inputs (than just depth). Maybe part-level segmentation?

• [ ] Use a pre-trained DeepSDF model:

  • If it is pre-trained on cars, then this should take the same route as point 1.
  • If it is pre-trained on chairs, then we have to devise a new set of experiments with 3D Future.

• [x] Substitute Pointnet encoder for Pointnet++ encoder

• [ ] Export the predicted colors along with the point coordinates. .ply is an ideal format.