Sunand Raghupathi

Adding the option to resize frames during DGP inference. This is important when, for example, inferring on videos that aren't in the training set. Also adding a script to easily...

Multi-level masking

3

Current approach just adds bit-conditioning to the latent vector, and determines which (simulated) ground truth mask to compute losses with.

Current OmniGAN tasks

1

- [ ] Hyperparameters optimization for masker and painter respectively - [ ] Try adding WD data (keep Unity as well) - [ ] Learning rate decaying - [ ]...

Masker Next Ideas

1

A list of short-term ideas to try for improving the masker: - [ ] Implement shared DA discriminator operating on mask + latent vector -- Sun - [ ] Use...

We've converged to a SPADE + MUNIT hybrid, but we need some tricks to improve. Here are the short term ideas to try: - [x] Gram Matrix as measure of...

Hardcoded domain adaptation classifier in MUNIT/SPADE

4

The domain adaptation classifier in the MUNIT/SPADE codebases (in "utils.py") doesn't work with arbitrary latent vectors. For example, changing the number of downsampling layers breaks the code

TODO: MaskGeneration

4

* [ ] Training on Unity Urban + Suburban *Méli* * [ ] Training on WD *Méli* * [ ] Training on heavier flooded images (Unity) * [ ] Conditioning...

Simulated data in flood generator doubles training time

4

The addition of simulated data (with domain adaptation) to the flood generator makes training significantly slower

Mask-conditioned encoder reduced water quality

2

Conditioning the encoder with the mask allows the network to not encode parts of the masked region (which increases destructiveness), but seems to result in worse water quality. Example: