No inpainted results generated on JPG about lama HOT 8 CLOSED

Hi! Sorry for the late reply!

It is by design that the input dataloader reads images only with a single variant of extension (either .jpg or .png). There are no sparticular strong reasons for that - just because it is simple and has been enough for evaluation purposes.

You can change the line dataset.img_suffix=.jpg to the prediction command here so the dataloader reads JPGs

from lama.

Please note that most our scripts use Hydra configuration system - so no need to change YAML files, for ad-hoc modifications use can just override any parameter via command line.

from lama.

Thanks for the reply @windj007.

Changing dataset.img_suffix=.jpg at the given location did not resolve the issue. I still get the message:

- Make val dataloader default from /home2/varungupta/lama/ts_images/
0it [00:00, ?it/s]

Please note that most our scripts use Hydra configuration system - so no need to change YAML files, for ad-hoc modifications use can just override any parameter via command line.

I was following the instructions mentioned in the readme, as mentioned under point 2.Prepare images and masks ---->
Specify image_suffix, e.g. .png or .jpg or _input.jpg in configs/prediction/default.yaml.

and thus updated the YAML to replace .png to .jpg. Am I missing something still?

from lama.

Hi! Is your question resolved? If not, could you please provide more details on what parameters do you use and how the data look like?

from lama.

No, the model does work as intended with .png, but not with .jpg files. I tried your suggested solution but that didn't work.
Also, you stated that

Please note that most our scripts use Hydra configuration system - so no need to change YAML files, for ad-hoc modifications use can just override any parameter via command line.

So, is the readme incorrect? (Point 2 of readme stated above). If so, kindly update it.

could you please provide more details on what parameters do you use and how the data look like?

I didn't change much parameters, only the YAML file chage mentioned in the readme. Further, I also tried your suggested update of dataset.img_suffix=.jpg. But the errors (mentioned in the issue) persisted.

Sample Image (in png format)

I managed around this issue by changing to .png format. So, this doesn't concern me as of now. Feel-free to close this issue, if you wish.

from lama.

Hi, everyone !

I've tested too, The same situation described by varungupta31 happens. For me it doesnt work editing the yaml file or passing as argument.

python bin\predict.py model.path=.\big-lama indir=.\jpg_images outdir=.\output dataset.img_suffix=.jpg

from lama.

Sorry for a long discussion and confusion. I forgot to highlight it from the beginning: dataset.img_suffix only changes how to obtain image filename from mask filename.

The algo is as follows:

1. obtain a list of mask files by pattern "*mask*.png", which cannot be changed from a config
2. for each mask file, obtain the corresponding image file by removing everything after _mask and adding the value of dataset.img_suffix

Thus, if you have masks in jpg, then changing configs will not help you - the script just will not find any masks in the input folder.

Please note that storing masks in JPG might be suboptimal - JPG is a lossless format and thus the mask after loading is not binary (only 0 and 1s) - it has a kind of gradient on mask boundary. The latter in its turn leads to severe artifacts.

Even if the masks are stored in PNG, the images can be stored in any other format, which PIL or cv2 can load - just set dataset.img_suffix accordingly. However, we did not test it with anything except PNG and JPG.

Takeaway: either store your masks in PNG or modify the code so it binarizes the masks after loading.

from lama.

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

from lama.