DIANNE annotation tool outputs xml annotations, this script was made to convert the dataset of the ROSE's chalange in COCO format

rose2COCO.py

import os
import xml.etree.ElementTree as ET
import json
import cv2
from PIL import Image

def rose2COCO(coco_annotations, root_dir, name_ds, name_cls):
    '''
    root_dir = base directory of the dataset 
    name_ds = name of the dataset
    name_cls = name of the class
    '''
    
    # Define the directory paths

    cls_dir = os.path.realpath(f'{root_dir}/COCO/{name_cls}')

    image_dir = 'src'
    annotation_dir = 'ref'
    years = ["2019","2020","2021"]
    img_data = os.path.realpath(f'{cls_dir}/data')
    

    # make the dir for the output
    if not os.path.exists(img_data): os.makedirs(img_data, exist_ok=True)


    # Initialize the category and annotation ID counter
    category_id = 0
    annotation_id = 1

    # Loop over the directories
    for year in years:
        year_dir = os.path.join(root_dir, year)
        image_set_dir = os.path.join(year_dir, image_dir, name_ds, name_cls)
        annotation_set_dir = os.path.join(year_dir, annotation_dir, name_ds, name_cls)
        for xml_ann in sorted(os.listdir(annotation_set_dir)):
            tree = ET.parse(os.path.join(annotation_set_dir, xml_ann))
            root = tree.getroot()

            # Extract the segmentation information
            segmentations = []
            category_ids = []
            for clipping in root.findall('.//clipping'):
                points = clipping.find('points')
                segmentation = []
                for point in points.findall('point'):
                    segmentation.append(int(point.attrib['x']))
                    segmentation.append(int(point.attrib['y']))
                segmentations.append(segmentation)
                # Extract the category information
                category = clipping.find('class').text
                if category == "unknown":
                    category_ids.append(2) # unknown are clasified as weed, we avoid to distinguish between known weeds
                else:
                    for cat in coco_annotations['categories']:
                        if category in cat['name']:
                            category_ids.append(cat['id'])


            # Compute the bounding boxes
            bboxes = []
            for segmentation in segmentations:
                xmin = min(segmentation[::2])
                xmax = max(segmentation[::2])
                ymin = min(segmentation[1::2])
                ymax = max(segmentation[1::2])
                bbox = [xmin, ymin, xmax - xmin, ymax - ymin]
                bboxes.append(bbox)

            # Extract the image information
            image_id = len(coco_annotations['images']) + 1
            file_name = xml_ann.replace('.xml', '.jpg')
            file_path = os.path.realpath(os.path.join(image_set_dir, file_name))
            if not os.path.exists(file_path):
                file_name = xml_ann.replace('.xml', '.png')
                file_path = os.path.realpath(os.path.join(image_set_dir, file_name))
                if os.path.exists(file_path):
                    img_png = Image.open(file_path)
                    file_path = file_path.replace('.png', '.jpg')
                    img_png.save(file_path)
                else: print(f"image not found {file_path}")
                    
            img = cv2.imread(file_path)
            height, width = img.shape[:2]
            image = {
                'id': image_id,
                'file_name': file_name,
                'height': height, 
                'width': width,  
            }
            coco_annotations['images'].append(image)

            target = f"{img_data}/{file_name}"        
            if not os.path.exists(target): 
                os.symlink( file_path, target)

            # Extract the annotation information
            for i in range(len(segmentations)):
                annotation_id = len(coco_annotations['annotations']) + 1
                annotation = {
                    'id': annotation_id,
                    'image_id': image_id,
                    'category_id': category_ids[i],
                    'segmentation': [segmentations[i]],
                    'bbox': bboxes[i],
                    'area': height*width,  
                    'iscrowd': 0,  
                }
                coco_annotations['annotations'].append(annotation)
                #annotation_id += 1
    return coco_annotations


if __name__ == "__main__":        
    # Initialize the COCO annotations
    coco_annotations = {
        'licenses': [],
        'info': {},
        "categories": [
          {
           "id": 1,
           "name": "crop",
           "supercategory": ""
          },
          {
           "id": 2,
           "name": "weed",
           "supercategory": ""
          }    
         ],
        'images': [],
        'annotations': [],
    }

  
    root_dir = './dataset/rose'
    ds_names = ['weedelec','bipbip','pead','roseau'] #, 
    cls_names = ['bean','maize']
    #cls_names = ['maize']
    
    
    for name_cls in cls_names:
        print(name_cls)
        coco_annotations['images'] = [] # reset annotations for each class
        coco_annotations['annotations'] = []
        label_files = []
        #print(len(coco_annotations['images']))
        for name_ds in ds_names:
            print("  " + name_ds)
            coco_annotations = rose2COCO(coco_annotations, root_dir, name_ds, name_cls)
            
        print("\n  " + str(len(coco_annotations['images'])))
        # Save the COCO annotations to a JSON file   
        cls_dir = os.path.realpath(f'{root_dir}/COCO/{name_cls}')
        label_file = os.path.realpath(f'{cls_dir}/labels.json')   
        with open(label_file, 'w') as f:
            json.dump(coco_annotations, f, indent=1)