inbalanced dataset?

[watermoon123]

Hi,

Thanks for you great work. I downloaded the data to check and found that it seems that most of the hair is relatively short. Is this because it's not easy to procedurally generate long hair?

[RaduAlexandru]

Hi @watermoon123

There is slightly more short hairs rather than long ones since there are some mode of variation (baldness, side-shave, etc) that are more likely to occur when hair is short so I wanted to sample them more often.
However there should be also a lot of long hairstyles in the dataset.

One other aspect is that the hairstyles base_idx are roughly ordered by length, so if you check the image folder the first few thousand samples is mostly short hair while the later samples are mostly long.
You can also use this info to resample and bias more towards a certain length if you need to retrain.