Could not reproduce the results demonstrated in the paper

[bananaman1983]

First off, I am very intrigued by the approach this project is taking. I think utilizing polarized photography is a valid way to tackle specular highlights since it should theoretically remove the specular reflections leaving only diffuse reflections.

However there seem to be some inconsistencies with the code or the dataset available on the github from the paper which causes differences in the resulting output.

I have managed to redirect the references to the dataset by modifying the train_specularitynet.py, to acquire checkpoints after around 100 epochs of training, but in some instances they do not seem to yield outputs equivalent to the results demonstrated by authors.

For instance, wider areas of specular reflections seem to get omitted. Take an example here:

I would be pleased to hear from the authors about some training details on the dataset and if the code on the github is the up-to-date version with which they used to produce the end result for the paper.

For the dataset, of the training set of 2210 scenes from the paper, substantial amount seem to be lost, leaving around 900 scenes at most.

As for the code, I have noticed some lines are commented out of spec.py where diffuse dataset is replaced by specular data, leaving calculated masks blank. I wonder if this was intentional. If so, I would be happy to know why this was necessary and your opinion on if reversing it would improve the results.

class SpecDataset(torch.utils.data.Dataset):
    def __init__(self, opt, datadir, dirA='spec', dirB='nospec',imgsize=None):
        super(SpecDataset, self).__init__()
        self.opt = opt
        self.datadir = datadir
        self.dirA = dirA
        self.dirB = dirB
        self.fnsA = sorted(os.listdir(join(datadir,dirA)))
        #self.fnsB = sorted(os.listdir(join(datadir,dirB))) 
        self.fnsB = self.fnsA
        self.imgsize = imgsize
        # np.random.seed(0)
        print('Load {} items in {} ...'.format(len(self.fnsA),datadir))

I could upload the fixed version in the pull request but I see that the license disclaimer is missing in this project and would feel relieved to know under what license the current code is distributed.

I admire the work you have published and would love to hear from you anytime soon.

[JovinLeong]

I'm facing the same issue - @bananaman1983 would it be possible to upload your fixed version? Thanks so much!!!

[bananaman1983]

I'm facing the same issue - @bananaman1983 would it be possible to upload your fixed version? Thanks so much!!!
Sorry. Without a clear declaration of what scope of license this project is built upon from the author, I don't think I am allowed to publicly post any form of code modified from the source.

[JovinLeong]

Understood, let's just wait on @jianweiguo then