Training Instance Segmentation model on torchvision.datasets.Cityscapes

Hello everyone,

I want to use the code in the Object Detection Finetuning Tutorial to train and test a Mask-RCNN model on torchvision.datasets.Cityscapes.

I am having trouble with the train_one_epoch function, similarly to what is happening in this issue 'train_one_epoch' gives error while using COCO annotations · Issue #1699 · pytorch/vision · GitHub. The solution is that targets should be a list of dictionaries, but I don’t know how to go from the type of target that we have in the Cityscapes class (PIL image) to a list of dictionaries like that.

Is there an easier way to train a model on Cityscapes without having to do this?
Thanks a lot,

It seems that the targets are expected in the “COCO format”, so this repository might be useful, which seems to convert Cityscapes to COCO.

1 Like