Perform baddly on validation set

[Choconuts]

I use the milkbox dataset to train the model and the training PSNR is about 29. However, the images synthesis in validation are nearly blank with a little noise, and the PSNR is extremely low. Is it overfitting?

I just run like this for the geometry model:

python train.py
--config
configs/milkbox_geometry.yaml
--datadir
data\milkbox_dataset

[zfkuang]

It might be overfitting. can you send me your log file?

Also, did you try the other two datasets?

[Choconuts]

version_2.zip not yet.

[zfkuang]

It looks like your initialization is a bit different from my side, what's your ubuntu version and did you use the conda command in the guide to install all the packages?

My initialization

Yours

[Choconuts]

I use windows. Does it matter so much? I mean initialization, since it can converge on training-set easily after all. How can I correct the initialization?

[zfkuang]

It might be, but there're some other things that looks odd: for example your camera loss is extremely high, indicating that the system didn't find a good camera pose either. So my guess is that there're some nuances between the packages on different platforms causing all the trouble. I would suggest using ubuntu, because all experiments are conducted on it. I can also test the model on windows later to see how to fix that.

[Choconuts]

OK, thank you.