Is it necessary to use the augmented Labelme features provided by you?

[PyJulie]

I tried to train from the original images and annotations downloaded from your homepage but obtain bad results using crowd layer and doctor net. But I achieved similar results using Majority-vote and DS labels. So I wonder if crowd layer still works when training from a dataset whose most samples lack annotations especially when there are so many annotators but most of them did not label all samples.