Questions about Classification of Custom Image Dataset

[jbawsume]

Hi . I read your paper and was really impressed by your results I was also looking into this topic. I had a question about how to modify the config /dataset.yml file I am interested in testing it on the State Farm distracted driver dataset which also has ten classes and was wondering if there are any requirements for file location and the .yml file I assume I need to create a State Farm.yml file

[XzwHan]

Hi @jbawsume, thank you for checking out our work. It is advised to create a new .yml file specifically for a new dataset, as well as a new .sh file under classification/training_scripts directory to run scripts for both training and evaluation. In the .yml file, you could modify data.dataroot argument to specify the directory where you plan to download and store the dataset; there are no requirements for file location as long as it's accessible. You shall also make sure that data.num_classes argument is set to $10$.

[jbawsume]

Sure thank you for taking the time to respond and one more question is there any requirement for dataset file structure. Currently the dataset is in the format dataset > train , test | train > with the ten sub folders representing the classes and test with only the test images no sub folders inside

[XzwHan]

You may check out how we load the image dataset with the get_dataset function in classification/utils.py, specifically Line 116-148.

[jbawsume]

Thank you for taking the time to explain that got it