Indoor photos would more likely be yellowish

[masterTW]

Hi,
Thanks for sharing data sets for SimpleCube++.
May I learn if you have trained any deep-learning models?

I recently used your data sets to train the auto-white balance model, and it turned out that
indoor photos would more likely be yellowish.
I am hoping to know if other people have also encountered similar issues, as I've checked the training program and made sure there's no mistake.

SimpleCube++ color temperature distribution shall be more realistic and the outcome of the trained model is not as expected.
May I ask what areas I shall check?

Thanks

[savfod-acmetric]

Hi!

We've trained some and I don't recall having such a problem. But it was long time ago and had it's own specifics, so I am not 100% sure (maybe @Ershoff recall smth like that?)

I would expect the issue to be related to the image processing pipeline as it is very tricky. In this case, checking the train images through the pipeline may help. (we have a demo pipeline in https://github.com/createcolor/nightimaging#data, but the data format may slightly differ)

For example, the issue may be related to the black level. Indoor images are usually less illuminated than in daylight, so any inaccuracy with black level subtraction (imho, almost inevitable) will have a greater impact on them.

(This could also be a problem if JPEG images are used for training, they are given for visualization purposes only. Because they may be rendered with a fixed white balance settings, as we don't aim to just reproduce some existing white balance algorithm. Hence, the JPEG indoor include more yellow images, as the indoor lighting is actually more yellowish).