How to perform an inference task with a model provided with trained parameters?

[YMX-zknu]

Hi author, I have downloaded the Cityscapes Trained Models and the file is as follows:

I would like to know how to set up the file to perform inference on the Cityspaces dataset with the model you provided.

thanks.

[twke18]

Hi,

You can use these python scripts for evaluation. We perform nearest neighbor retrieval to evaluate the trained models.

[YMX-zknu]

Thanks, but I'm having another problem, as shown below.

I tried to run the file "generate_pmi_owt_ucm.m" but I did not find the files corresponding to variables 'matDir', 'segDir', 'imgNames' in the program files you provided.

[twke18]

Sorry that I could not give you exact examples, as I lost the access to my code repo. I will instead explain what these variables are:

1. matDir is a cache directory which saves the edge maps computed by PMI. The code will create the directory.
2. segDir is the directory which saves all the segmentation masks by thresholding the edge maps. The code will also create the directory.
3. imgNames is a list of image file names which can be found under rootDir (code). Originally, imgNames are read from these lines. You can modify it in your way to read the image names accordingly.

[YMX-zknu]

Thank you, there are some errors when executing the inferance.py file, I hope you can answer them, as described below:

I have constructed the Cityscapes dataset as per your requirement as shown below:

However, while executing the inferance.py file, the following error occurs:

The displayed "gtFineInstId" seems to require instance segmentation images, but this file is not present in the built Cityscapes dataset. What is the reason for this? Thank you very much!

[twke18]

Hi, the files are the ground-truth instance masks of Cityscapes. They are not used for inference at all. You can either replace them with some dummy files, or download from here.

[YMX-zknu]

Thank you very much, I downloaded the gtFineInstId.zip file you provided and am currently able to run the entire program.

I would like to train a new model based on my dataset to replace the pre-trained model you provided. Currently encountering two problems:

1. As shown below, the variable imgNames in the generate_pmi_owt_ucm.m file reads from the cityscapes_train.txt file. Is the cityscapes_train.txt file the path to the images within all the sub-files of the leftImg8bit/train file in the established Cityscapes dataset?
   
   Also, what is the purpose of these '*.txt' files (shown below) under the datasets/cityscapes path in your project?
   

2. I want to modify my dataset to the layout of the Cityscapes dataset you provided, but I don't know how the contents of the folder gtFineId are generated? Is it generated based on the gtFine_trainvaltest.zip file (shown below) on the Cityscapes website? Is it convenient for you to provide the corresponding processing code?
   

Thank you very much for your time and help and I look forward to hearing from you.