Very low scores for provided pretrained model

[xvbw]

Hi, I tried to run 'eval_all.m' code to see the performance of pretrained model for Cityscapes dataset However, even if I followed the instruction, the performance is very very low(See as below) This is the result of 5 images of Cityscapes. As you can see, the performance is really bad. I would like to ask what is the problem. Below is the code. I just changed paths for database.

I suspect label png files I used are wrong. I used png files included in Cityscapes dataset. The file format is like 'frankfurt_000000_000294_gtFine_labelIds.png' ... and so on.

Thanks!

[jmydurant]

Hi, @xvbw . I face the same problem recently. Cityscapes dataset annotation has two kind label : "id", and "trainid". Unfortunately, gtFine_labelIds.png uses id and the image in the list uses another one. To get the right images, you should clone the official code [https://github.com/mcordts/cityscapesScripts]. then put the code to the root path of the dataset. then execute

python preparation/createTrainIdLabelImgs.py

Good luck!!

[xvbw]

@jmydurant Thanks! I've checked out that site too before and guessed it would be a solution. That worked. Thanks again