How to convert yolo annotations to coco format. Json?

Question

I want to convert my labels in yolo format to coco format I have tried https://github.com/Taeyoung96/Yolo-to-COCO-format-converter And Pylabel They all have a bugs.

I want to train on detectron 2 but it fails to load the dataset because of the wrong json file.

Lotti

Could you elaborate on the error you get? Also, provide file examples. – Connive

Answer 1

You can use the globox package (disclaimer: I'm the author). This should ressemble something like:

from pathlib import Path

import globox


def main() -> None:
  path = Path("/path/to/annotations/")  # Where the .txt files are
  save_file = Path("coco.json")

  annotations = globox.AnnotationSet.from_yolo(gts_path)
  annotations.save_coco(save_file)


if __name__ == "__main__":
    main()

If you need to map the labels using a .name file:

names_file = Path("/path/to/classes.names")
names = AnnotationSet.parse_names_file(names_file)
annotations.map_labels(names)

you can check the documentation for other formats (Yolo Darknet, Yolo v8, etc.).

Connive

Hello thanks for your help but i get this error File "C:\Users\Nima\Desktop\ObjectDetectionEval-master\ObjectDetectionEval\boundingbox.py", line 186, in <genexpr> coords = (float(c) for c in coords) ValueError: could not convert string to float: '' – Lotti

An annotation file may be malformed, could you investigate which one is the culprit and post it here please? – Connive

I tracked down the bug my labels in the first line have an space in end and that makes your code not to work example:0 0.5375 0.8875 0.10625 0.212** this * characters are spaces . is there any way to strip end of the lines in your code – Lotti

I will send an example of the label file your GitHub repository issues – Lotti

Yes, line 175 of boundingbox.py you can add a .strip() after the .split(), this should solve the issue. I'll fix the bug in the repo. – Connive

before*, not after – Connive

There is no bug in your repo it was bug in my dataset . i fixed my dataset with script and your code worked like charm – Lotti

Sure, but fixing it makes my repo more robust to these kind of dataset bugs. You can accept my answer if it solved your issue, thanks! – Connive

There is little problem: there is no method named Annotation.parse_names_file(names_file) is my repo outdated? and I want to map categories from 1 instead of zero is there anyway for this – Lotti

The function name is parse_names_file(). You can use the map_labels() method to map to different labels. – Connive