labelme2coco.py bounding box problems

[Darshan2701]

Hi fellow researchers,
I have encountered a problem and need a bit of clarification and help. Kindly read the following..
Steps to try:

1. Annotate the same objects (2 or more) in an image by using a rectangle (Not polygons)
2. Save the file as json.
3. Convert the existing json to coco format using the labelme2coco.py file.
4. Compare the jsons.

I see that there is only one bbox field in the coco converted json though there are more than one bbox in the image. But the number of segmentations in the coco converted json matches with the non coco json file.

Is this an issue with the code? Or is this intended to be like this?
Kindly clarify.

[wkentaro]

Or is this intended to be like this?

Yes. Because current labelme2coco.py assumes the annotation is a segment instead of a bounding box.
The way of solving this is

• create new labelme2coco_bbox.py
• or
• name different label to different bbox (e.g., person-1, person-2)

[JennyLouise]

I replaced lines 109 to 125 in the original labelme2coco.py file with the following, renaming labels to person-1, person-2, etc, before adding them to the coco annotations file.

        masks = {}                                     # for area
        segmentations = collections.defaultdict(list)  # for segmentation
        labels_indices = {}

        for shape in label_data['shapes']:
            points = shape['points']
            label = shape['label']
            shape_type = shape.get('shape_type', None)
            mask = labelme.utils.shape_to_mask(
                img.shape[:2], points, shape_type
            )

            if label in masks:
                labels_indices[label] += 1
                label=label+"-"+str(labels_indices[label])
                
            else:
                labels_indices[label]=0

            masks[label] = mask