Detectron2 (only) semantic segmantation annotation format.

[rozeappletree]

I want to train model for semantic segmentation only. with

• Input: single channel image
• Target: binary sematic segmenation mask.

To build the dataset, the following keys are REQUIRED. But I don't have them because I am doing ONLY semantic segmantation right!

• bbox
• bbox_mode
• segmentation

Currently I am using this code to build my dataset dicts

""" create binary mask annotations for detectron2 """
import cv2
import json


def convert_to_d2sseg(img, msk):
    imgmat = cv2.imread(str(img), cv2.IMREAD_GRAYSCALE)    
    return {
        'file_name': str(img),
        'width': imgmat.shape[0],
        'height': imgmat.shape[1],
        'image_id': msk.stem,
        'annotations': [{
                'bbox': [], 'bbox_mode': 0,
                'category_id': 0,
                'sem_seg_file_name': str(msk)
            }
        ]

    }


TRAIN_DIR_LOCAL = Path('/kaggle/working/data/train') 
VALID_DIR_LOCAL = Path('/kaggle/working/data/valid')
TEST_DIR_LOCAL = Path('/kaggle/working/data/test')


for ddir in [TRAIN_DIR_LOCAL, VALID_DIR_LOCAL, TEST_DIR_LOCAL]:
    imgs = sorted(list(ddir.glob('*-image.png')))
    msks = sorted(list(ddir.glob('*-mask.png')))
    for img, msk in tqdm(zip(imgs, msks), total=len(imgs)):
        dic = convert_to_d2sseg(img, msk)
        fp = ddir / f'{msk.stem}.d2.json'
        with open(str(fp), 'w') as f:
            json.dump(dic, f)

[github-actions[bot]]

You've chosen to report an unexpected problem or bug. Unless you already know the root cause of it, please include details about it by filling the issue template.
The following information is missing: "Instructions To Reproduce the Issue and Full Logs"; "Your Environment";

[rozeappletree]

• Instructions To Reproduce the Issue and Full Logs: not needed
• Your Environment: kaggle

[rozeappletree]

I am ready to send a PR to docs and current colab as it is not present for custom data. I just need to know what is dictionary format it we are doing semantic segmentation only

[hiroto01]

@INF800 if you go on reading the docs->(https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html)
youll get a better idea it's stated to some extent what needs to be done there..i'll be happy to help if you need any..

[rozeappletree]

Can you share the dictionary format for smantic segmentation.

[ppwwyyxx]

I think it's already pretty clear from the link given above.

Moving to discussions since I believe it is not a documentation issue.

[hiroto01]

@INF800 where did you provided colab link?

[rozeappletree]

Not provided yet. It is in my private kaggle kernel. Will move it to colab soon

[mjb-oz]

@INF800 Did you ever move your PointRend semantic segmentation code to Colab or somewhere else public?
I'd love to have a look if you have a link to it.

[sergiev]

@INF800 @mjb-oz hi there, i'm looking for working example of semantic segmentation, could you share any?

[mjb-oz]

Hi @sergiev, I don't have any working semantic segmentation Dectectron2 code. Would love to see some though.