hello,when i want to use SceneParsing how i process individual components? about cascadepsp

[hkchengrex]

It can be done by appending the --ade flag to eval.py. See here.

[zhujiesuper]

I add --ade but only have one class mask ，the other class i can not find

[zhujiesuper]

if I use the wrong data ，I use ADE_EncNet_SS_ours for eval.py

[hkchengrex]

You need to use the "Pre-processed 'split' input" here. We pre-process the scene parsing data into segments for easier handling.

[zhujiesuper]

okok，the eval.py run correctly,but the eval_post_ade broken: new_class_i[c] += new_i
IndexError: list index out of range
can you tell me such like --mask_dir XXX ,---seg_dir YYY,,,,,the XXX,andYYY and ZZZ I do not konw the data is what in detail ,can you support this to me?

[hkchengrex]

They are mentioned here.

[zhujiesuper]

sorry ，i see this,but i do not solve becouse i do not not what data to put the correct local. can you give me a clarity note. such like --mask_dir XXX ,---seg_dir YYY,,,,,the XXX,andYYY and ZZZ,and XXX use what data from the onedrive,like ADE_PSPNet_MS_ours

[hkchengrex]

mask_dir -- The output folder you generated with eval.py
seg_dir -- Download from the "segmentation inputs" column
gt_dir -- The "annotations" folder downloaded from the official ADE20K website
split_dir -- The input folder you used in eval.py

[zhujiesuper]

thank you! I run this well ,but the "segmentation inputs" column whether is the original pspnet output?
and Break down the scene parse into individual components using the method described in the paper,,,,,do this have code which not provide?

[hkchengrex]

The code is in the repo, in scripts/ade20K. They are not documented and quite messy though.

[zhujiesuper]

okok, the "segmentation inputs" column whether is the original pspnet output?

[zhujiesuper]

and can you provide scripts/ade20K run step,which i run first and next?

[hkchengrex]

1. If you are using the PSPNet one then yes.
2. They are only for your reference and are not that "plug-and-play". Look at ade_expand_inst.py for starters. This is because each source (PSPNet, deeplab, etc.) produce their output differently. We need some manual labour to "unify" the format.

[zhujiesuper]

thank you very much.