Influence: Amount of training data and data augmentation via Detectron2 about lama HOT 3 CLOSED

Hi! Sorry for the late reply!

How much did the larger training data (4.5M images from the Places-Challenge dataset) contribute to the improved quality of Big LaMa-Fourier?

Larger data helps, but significantly less than larger model size and other training tricks (SegmPL, large masks)

Do you think that similar results could have also been achieved for Big LaMa-Fourier with less data?

Less data = less quality, but not dramatically - reducing model size or removing SegmPL or using smaller training masks would hurt more

How much did the training and the inference benefit from data augmentation using segmentation masks from Detectron2?

We do not use segmentation masks from Detectron for training. We tried it in the very beginning of the project, but faced technical issues (slow, gpu memory consumption, cuda-reinitialization limitation) - so we could not use segmentation-based mask generation effectively during training. It is there just because we forgot to remove it when preparing a public code release. Please note that segm_proba: 0 in all the data configs.

from lama.

In our experience, mask widths matter much more than mask shapes

from lama.

Hi Roman,
awesome!! Thank you for sharing your experience! This really helps a lot and confirms my preliminary findings :)

from lama.