training on novel dataset

[Harry Carey]

Hey, great paper! any advice on how to train IAST on a new dataset?

[Wenqi Tang]

Sorry for the late reply, thank you for your attention to our paper.

For training IAST on a new dataset, please follow the following steps:

1. Generate a JSON file where each item should contain image_name and mask_name, which indicate the relative path of image and label. For more details, please see the JSON files in IAST/data.
2. Implement a dataset class that inherits the BaseDataset, just like the MsCityscapesDataset, and overwrite the transform_mask(self, label) for specific label transformation and aug(self, image, label) for the augmentation. Last but not least, please register your dataset with a customized name by using @DATASET.register("your_dataset_name").
   

   IAST/code/sseg/datasets/loader/cityscapes_dataset.py

   Lines 17 to 25 in 24db991

    @DATASET.register("MsCityscapesDataset") 
    class MsCityscapesDataset(BaseDataset): 
        # overwrite 
        def transform_mask(self, label): 
            return label 
         
        # overwrite 
        def aug(self, image, label): 
            return crop_aug(image, label, 512, 1024, min_max_height=(341, 1000), w2h_ratio=2) 

I hope this can help you, I close this issue, if you have any other questions, please comment and I will reopen it.