search

Atif-Anwer / SHMGAN

Joint network for specular highlight detection and adversarial generation of specular-free images trained with polarimetric data
https://authors.elsevier.com/a/1hqRr3INukOUwT
3 stars 0 forks source link

Dataloader misaligned with PSD dataset #2

Open JovinLeong opened 1 month ago

JovinLeong commented 1 month ago

Hi, I'm trying to train SHMGAN on the PSD datasets with main.py, however, it seems like the PSD dataloaders seems to expect a different directory structure with how the PSD dataset that was linked in the README is structured.

Data loader:

image

Actual PSD structure:

image

Could I clarify which dataset I should be using here to reproduce the results in the paper and how best to go about training SHMGAN on both PSD and SHIQ?

Atif-Anwer commented 1 month ago

The SHMGAN dataset commented out is our in-house dataset that I acquired using the polar cam (and that has not been released); and that code is there for legacy purposes.

The PSD dataset as far as I remember was only organized into test-train-val groups. For training, the dataset was re-organized into subfolders for each angle. In the PSD_Train/aligned folder, their images are named from 0-12 for a difference of 30 degree angles. Out of these, we only used the 4 as shown in the images above. But the dataset was reorganized into four separate folders, 1 folder for one angle, so that the data loading was easier. I should have the reorganizing script somewhere but ill have to search a bit. But i hope this explains the issue.