Can I train it on the CoCo Panoptic dataset?

[squashking]

Sorry for the newbie question. I see there is train_coco.sh file, I am just wondering if it can be used to train the Coco Panoptic dataset. If so, how should I organize the train data (i.e. the directory structure of the dataset)? Thanks a lot!

[RubenS02]

I have not seen the blocks (instance segmentation head, instance segmentation losses) needed for the instance segmentation task but you could maybe add them yourself

[squashking]

@RubenS02 Thanks. That's too challenging for me since I'm a beginner. I'd go for semantic segmentation instead. Can you please tell me how to organize the train data? I mean, using the Coco 2017 train dataset. Thank you!

[RubenS02]

@squashking I have not trained with COCO, but everything you need is in dataloaders/datasets/coco.py

[kungkook]

I am also a novice, can I train my own data set? I also do semantic segmentation, I need to manually make the data set, how should I use this network training, and what should I modify?

[faheuer]

Please dont nectropost. Making your own dataset is a LOT of work. You can look into tools like labelme for that. Then, use a dataset format like coco style, and modify the coco.py dataset to suit your own.