Random Mask Generation about lama

[windj007]

When I needed to quickly draw a mask, I used GIMP several times - just load an image, create another transparent layer, draw a mask there and save it to a separate file. Note that all masks must be strictly binary before feeding them to LaMa - thus if you make them with GIMP, you need a separate binarization step (that's pretty easy to do with cv2 and plain numpy).

[windj007]

I'm not sure that I understand the question right... For better quality, the mask should cover an object and maybe a bit more - even tiny unmasked parts may result in artifacts. However, mask must not follow object boundaries - LaMa does not depend on its shape - it sees only unmasked areas and tries to regenerate the image based only on the available information. LaMa does not refine mask shape in any way. Does this answer your question? If not, could you please clarify it?

[windj007]

Hi! Could you please describe your problem in more detail? What link is broken?

If you have an environment set up, you can use script

bin/gen_mask_dataset.py configs/data_gen/<config_name>.yaml \
    <dir with input images> <where to store image crops and masks>

Alternatively, there is a convenience wrapper via docker - please refer to the readme.

[ZeeRizvee]

Thank you for the help! Is there also an option to draw the mask on top of the image, like the one which is used in Google Colab demo?

[windj007]

No, we do not provide any UI except those in Colab.

[windj007]

You can try third-party apps, e.g. lama-cleaner.

[ZeeRizvee]

Thank you so much for the help and advice! Much appreciated :)

[ZeeRizvee]

@windj007 just another small query related to masks. As it can be seen in the Colab demo and lama-cleaner implementation, does the original LaMa repository also have the ability to recognize and stick with the main features in the original image even if the created mask does not perfectly align? If so, can you please point me towards the script which is performing this function?

[windj007]

Note about mask binarization: predict.py actually does binarization - I forgot about this line. But anyway, as stated in another issue, having smooth masks is meaningless and it's better to control it manually.

[ZeeRizvee]

My apologies for not being able to ask the question clearly. But you have answered quite elaborately and my confusion is clear now. Thank you for the clarification!

[windj007]

You're welcome!