Open haseeb33 opened 1 month ago
This is feasible, but I'm not sure if it's very simple.
We first use the datasets/prepare_ade20k_sem_seg.py
, datasets/prepare_ade20k_pan_seg.py
, and datasets/prepare_ade20k_ins_seg.py
scripts provided by Mask2Former to convert ADE20K annotations into the COCO format, and then use continual/memory_generator/memory_selection_pan.py
, continual/memory_generator/memory_selection_seg.py
, and continual/memory_generator/memory_selection_inst.py
for splitting the training data for each step.
I don't remember if Mask2Former provides code to convert the CityScapes dataset into COCO format. If it does, you should be able to use that code to first convert the CityScapes dataset into COCO format, then make slight modifications to continual/memory_generator/memory_selection_pan.py
, continual/memory_generator/memory_selection_seg.py
, and continual/memory_generator/memory_selection_inst.py
to split the dataset for each step. If you're using replay samples, continual/memory_generator/memory_selection_pan.py
, continual/memory_generator/memory_selection_sem.py
, and continual/memory_generator/memory_selection_inst.py
will also need corresponding changes. These modifications are not complicated; they mainly relate to the total number of classes, the range of classes in each step, and the total number of steps. You should be able to understand it after taking a quick look.
Can we easily modify the code to run BalConpas for CITYSCAPES dataset? Is yes, Can you please list down the steps required to modify the code?