search

zhang-tao-whu / DVIS

DVIS: Decoupled Video Instance Segmentation Framework
MIT License
127 stars 6 forks source link

where coco2ytvis2019_train.json? #27

Open zhending111 opened 10 months ago

zhending111 commented 10 months ago

coco2ytvis2019_train.json

zhang-tao-whu commented 10 months ago

You need to run convert_coco2ytvis.py to generate these files.

zhending111 commented 10 months ago

but I run it,got error:No such file or directory: coco/annotations/coco2ytvis2019_train.json' I find in code:os.path.join(_root, "coco/annotations/coco2ytvis2019_train.json"), "COCO to YTVIS 2019:"), should i prepare it?Thanks!

zhang-tao-whu commented 10 months ago

You need to make sure that the coco/annotations/ folder exists and prepare the coco and ytvis19 original datasets.

zhending111 commented 10 months ago

I don’t need to prepare the json file in the coco folder, right?your work is great，I will quote！

zhang-tao-whu commented 10 months ago

Thank you very much! Yes. coco2ytvis2019_train.json is automatically generated by convert_coco2ytvis.py based on the annotations from both COCO and ytvis19 datasets.

fbragman commented 8 months ago

Hi @zhang-tao-whu

Just a quick question on annotations - in ytvis_2019, I have annotations for validation in valid/Annotations but my valid.json file downloaded from the server does not have annotations any. I cannot compute AP/AR locally. So the following is False despite valid/Annotations existing

self._do_evaluation = "annotations" in self._ytvis_api.dataset

Is there a way to recreate the json file to also point to the annotations? I guess these were recently released since prior you could only validate on CodeLab?

Also, I am just wondering for code development and prototyping of your model, what is the best practice since validation of OVIS/YTVIS2019-2021 is only performed on the CodeLab server.

Do you create a train/val split from the train.json file? How did you evaluate model performance to ensure it was working?