Preparing data for Amodal Segmentation

[jitaxis]

Hi,

I currently have images annotated in the COCO format, which from my understanding is correct for the conversion used by BCNet?
When I try to run the line from process.sh, I get the following error:
File "/content/BCNet/detectron2/data/datasets/process_dataset.py", line 17, in
from .. import MetadataCatalog, DatasetCatalog
ImportError: attempted relative import with no known parent package

Also, I am trying to use BCNet on Google Colab.

Please let me know what I am doing or have done wrong and any further guidance is greatly appreciated.

[lkeab]

You could refer to the discussion here, which may be related with the python versions.

[jitaxis]

I think I realised what I did wrong. I am using @Shadercloud version from this issue #81 (comment)_

Therefore I didn't build the main BCNet repo, which means the packages were not found.

But when trying to run the same script from @Shadercloud version, I get the following error.

File "/content/detectron2/detectron2/data/datasets/process_dataset.py", line 575, in
assert sys.argv[3] in DatasetCatalog.list()
AssertionError

Which makes me think I am somehow passing my arguments incorrectly. An example of the line I am executing is below:

!python -m detectron2.data.datasets.process_dataset "/content/drive/Shareddrives/Collected Data/training_anno_images.json" "/content/drive/Shareddrives/Folder For Datasets/datasets/training_image_folder" "New_Name"

Side question:
For Amodal segmentation, does my training dataset have to be all not occluded images?

[lkeab]

For the fist question, the "new_name" should be registered in the config file, you could follow the way of "coco_2017_train".
For the side question, no.

[jitaxis]

So I have tried to register my coco annotated dataset using the following code:

train_json_annot_path = "/content/drive/Shareddrives/Collected Data/training_anno_images.json"
train_images_path = "/content/drive/Shareddrives/Folder For Datasets/datasets/training_image_folder"
train_dataset_name = "Name_train"

register_coco_instances(name = train_dataset_name,metadata={},json_file=train_json_annot_path,image_root=train_images_path)

!python -m detectron2.data.datasets.process_dataset train_json_annot_path train_images_path train_dataset_name

I still get the output:

File "/content/detectron2/detectron2/data/datasets/process_dataset.py", line 575, in
assert sys.argv[3] in DatasetCatalog.list()
AssertionError

Is this what you mean by "follow the way of "coco_2017_train"". If not could you please direct me more to solve this issue.
Thank you for all the help.