search

XiaohangZhan / deocclusion

Code for our CVPR 2020 work.
Apache License 2.0
794 stars 104 forks source link

Order matrix not correct even after adjusting 'th' and 'dilate_kernel' #32

Closed jiunyen-ching closed 3 years ago

jiunyen-ching commented 3 years ago

Hi, I was testing out the demo codes on this RGB image image

and its instance mask image

I slowly increased the 'th' and 'dilate_kernel' from (0.1, 5), (0.3, 7), (0.5, 9), (0.7, 11) respectively as advised in #14 but the order matrix is still not correct.

image

From the order matrix, Instance 3 (table) is not considered as an occluder to Instance 5 (bench). May I ask for some advice as to how to improve the situation?

XiaohangZhan commented 3 years ago

Why are there margins between objects in the segmentation map? These may severely affect the results. Dilation process can mitigate it but not thoroughly . You might have to deal with the margins before testing it.

jiunyen-ching commented 3 years ago

Unfortunately the margins are due to the NYU annotations. I will dilate the masks and try to infer their order again. Will update here if it works

jiunyen-ching commented 3 years ago

The masks not tightly fitting each other is indeed quite troublesome to handle. Only certain cases where the masks are tightly fitting, it works well.