How were the synthetic scribbles generated during training?

[aL3x-O-o-Hung]

So I generated some fake scribbles on our dataset using the ground truth labels but the pretrained model cannot predict anything meaningful while using the generate clicks and bounding boxes are fine. I am wondering how the scribbles are generated and how the generated scribbles can shift the input distribution so much that it completely confuse the network.

[halleewong]

Can you show some examples?

The scribbles were generated using the code in scribbleprompt/scribbles.py during training.

[aL3x-O-o-Hung]

Hi!

Thanks for the reply. Please find these examples attached.

[halleewong]

Those scribbles look quite thick, relative to the size of the images. The random deformations in our scribble generation code vary the scribble thickness a bit but the training scribbles are rarely as wide and as sharp as these examples. I suspect the predictions will be much better with thinner scribbles and/or lightly blurring the scribble mask.