Source code for megadetector.data_management.labelme_to_coco

"""

labelme_to_coco.py

Converts a folder of labelme-formatted .json files to COCO.

"""

#%% Constants and imports

import ast
import os
import sys
import json
import argparse

from multiprocessing.pool import Pool, ThreadPool
from functools import partial
from tqdm import tqdm

from megadetector.utils import path_utils
from megadetector.visualization.visualization_utils import open_image


#%% Support functions

def _add_category(category_name,category_name_to_id,candidate_category_id=0):
    """
    Adds the category [category_name] to the dict [category_name_to_id], by default
    using the next available integer index.
    """

    if category_name in category_name_to_id:
        return category_name_to_id[category_name]
    while candidate_category_id in category_name_to_id.values():
        candidate_category_id += 1
    category_name_to_id[category_name] = candidate_category_id
    return candidate_category_id


def _process_labelme_file(image_fn_relative,
                          input_folder,
                          use_folders_as_labels,
                          no_json_handling,
                          validate_image_sizes,
                          category_name_to_id,
                          allow_new_categories=True):
    """
    Internal function for processing each image; this support function facilitates parallelization.
    """

    result = {}
    result['im'] = None
    result['annotations_this_image'] = None
    result['status'] = None

    image_fn_abs = os.path.join(input_folder,image_fn_relative)
    json_fn_abs = os.path.splitext(image_fn_abs)[0] + '.json'

    im = {}
    im['id'] = image_fn_relative
    im['file_name'] = image_fn_relative

    # If there's no .json file for this image...
    if not os.path.isfile(json_fn_abs):

        # Either skip it...
        if no_json_handling == 'skip':
            print('Skipping image {} (no .json file)'.format(image_fn_relative))
            result['status'] = 'skipped (no .json file)'
            return result

        # ...or error
        elif no_json_handling == 'error':
            raise ValueError('Image file {} has no corresponding .json file'.format(
                image_fn_relative))

        # ...or treat it as empty.
        elif no_json_handling == 'empty':
            try:
                pil_im = open_image(image_fn_abs)
            except Exception:
                print('Warning: error opening image {}, skipping'.format(image_fn_abs))
                result['status'] = 'image load error'
                return result
            im['width'] = pil_im.width
            im['height'] = pil_im.height

            # Just in case we need to differentiate between "no .json file" and "a .json file with no annotations"
            im['no_labelme_json'] = True
            shapes = []
        else:
            raise ValueError('Unrecognized specifier {} for handling images with no .json files'.format(
                no_json_handling))

    # If we found a .json file for this image...
    else:

        # Read the .json file
        with open(json_fn_abs,'r') as f:
            labelme_data = json.load(f)
        im['width'] = labelme_data['imageWidth']
        im['height'] = labelme_data['imageHeight']

        if validate_image_sizes:
            try:
                pil_im = open_image(image_fn_abs)
            except Exception:
                print('Warning: error opening image {} for size validation, skipping'.format(image_fn_abs))
                result['status'] = 'skipped (size validation error)'
                return result
            if not (im['width'] == pil_im.width and im['height'] == pil_im.height):
                print('Warning: image size validation error for file {}'.format(image_fn_relative))
                im['width'] = pil_im.width
                im['height'] = pil_im.height
                im['labelme_width'] = labelme_data['imageWidth']
                im['labelme_height'] = labelme_data['imageHeight']

        shapes = labelme_data['shapes']

        if ('flags' in labelme_data) and (len(labelme_data['flags']) > 0):
            im['flags'] = labelme_data['flags']

    annotations_this_image = []

    if len(shapes) == 0:

        if allow_new_categories:
            category_id = _add_category('empty',category_name_to_id)
        else:
            assert 'empty' in category_name_to_id
            category_id = category_name_to_id['empty']

        ann = {}
        ann['id'] = im['id'] + '_ann'
        ann['image_id'] = im['id']
        ann['category_id'] = category_id
        ann['sequence_level_annotation'] = False
        annotations_this_image.append(ann)

    else:

        for i_shape,shape in enumerate(shapes):

            if shape['shape_type'] not in ('rectangle','polygon'):
                print('Skipping an annotation of type {} in {}'.format(
                    shape['shape_type'],image_fn_relative))
                continue

            ann = {}
            annotations_this_image.append(ann)

            if use_folders_as_labels:
                category_name = os.path.basename(os.path.dirname(image_fn_abs))
            else:

                label = shape['label'].strip()
                assert len(label) > 0, 'Illegal label for file {}'.format(image_fn_relative)

                category_name = label

                # When multiple fields are available, labelme populates a dictionary (or at least
                # I've seen this happen once)
                if label[0] == '{' and label[-1] == '}' and 'name' in label:

                    try:
                        label_dict = ast.literal_eval(label)
                        category_name = label_dict['name']
                        # Copy all other fields directly to the output annotation
                        for k in label_dict:
                            if k != 'name':
                                ann[k] = label_dict[k]
                    except Exception:
                        print('Warning: failed to parse what looked like a dictionary for {}'.format(image_fn_relative))

                        # This is a no-op, but including it for clarity
                        category_name = label

                # ...if we might have to parse a dict for this label

            if allow_new_categories:
                category_id = _add_category(category_name,category_name_to_id)
            else:
                assert category_name in category_name_to_id
                category_id = category_name_to_id[category_name]

            points = shape['points']

            # Populate non-geometry fields
            ann['id'] = ann['id'] = im['id'] + '_ann_' + str(i_shape).zfill(3)
            ann['image_id'] = im['id']
            ann['category_id'] = category_id
            ann['sequence_level_annotation'] = False

            # Populate geometry
            if shape['shape_type'] == 'rectangle':

                if len(points) != 2:
                    print('Warning: illegal rectangle with {} points for {}'.format(
                        len(points),image_fn_relative))
                    continue

                p0 = points[0]
                p1 = points[1]
                x0 = min(p0[0],p1[0])
                x1 = max(p0[0],p1[0])
                y0 = min(p0[1],p1[1])
                y1 = max(p0[1],p1[1])

                bbox = [x0,y0,abs(x1-x0),abs(y1-y0)]
                ann['bbox'] = bbox

            else:

                assert shape['shape_type'] == 'polygon'

                if len(points) < 3:
                    print('Warning: illegal polygon with {} points for {}'.format(
                        len(points),image_fn_relative))
                    continue

                # COCO segmentation is a list of polygons, each a flat list of coordinates
                segmentation = []
                for pt in points:
                    segmentation.append(pt[0])
                    segmentation.append(pt[1])

                # Compute the axis-aligned bounding box
                x_coords = [pt[0] for pt in points]
                y_coords = [pt[1] for pt in points]
                x0 = min(x_coords)
                y0 = min(y_coords)
                bbox_w = max(x_coords) - x0
                bbox_h = max(y_coords) - y0

                bbox = [x0,y0,bbox_w,bbox_h]
                ann['bbox'] = bbox
                ann['segmentation'] = [segmentation]

            # ...if this is a rectangle/polygon

        # ...for each shape

    # ...if we do/don't have shapes for this image

    result['im'] = im
    result['annotations_this_image'] = annotations_this_image

    return result

# ...def _process_labelme_file(...)


#%% Main function

[docs] def labelme_to_coco(input_folder, output_file=None, category_id_to_category_name=None, empty_category_name='empty', empty_category_id=None, info_struct=None, relative_paths_to_include=None, relative_paths_to_exclude=None, use_folders_as_labels=False, recursive=True, no_json_handling='skip', validate_image_sizes=True, max_workers=1, use_threads=True): """ Finds all images in [input_folder] that have corresponding .json files, and converts to a COCO .json file. Currently supports bounding box and polygon annotations, as well as image-level flags (i.e., does not support point annotations). Polygon annotations produce COCO annotations with a "segmentation" field and an axis-aligned bounding box. Labelme's image-level flags don't quite fit the COCO annotations format, so they are attached to image objects, rather than annotation objects. If output_file is None, just returns the resulting dict, does not write to file. if use_folders_as_labels is False (default), the output labels come from the labelme .json files. If use_folders_as_labels is True, the lowest-level folder name containing each .json file will determine the output label. E.g., if use_folders_as_labels is True, and the folder contains: images/train/lion/image0001.json ...all boxes in image0001.json will be given the label "lion", regardless of the labels in the file. Empty images in the "lion" folder will still be given the label "empty" (or [empty_category_name]). Args: input_folder (str): input folder to search for images and Labelme .json files output_file (str, optional): output file to which we should write COCO-formatted data; if None this function just returns the COCO-formatted dict category_id_to_category_name (dict, optional): dict mapping category IDs to category names; really used to map Labelme category names to COCO category IDs. IDs will be auto-generated if this is None. empty_category_name (str, optional): if images are present without boxes, the category name we should use for whole-image (and not-very-COCO-like) empty categories. empty_category_id (int, optional): category ID to use for the not-very-COCO-like "empty" category; also see the no_json_handling parameter. info_struct (dict, optional): dict to stash in the "info" field of the resulting COCO dict relative_paths_to_include (list, optional): allowlist of relative paths to include in the COCO dict; there's no reason to specify this along with relative_paths_to_exclude. relative_paths_to_exclude (list, optional): blocklist of relative paths to exclude from the COCO dict; there's no reason to specify this along with relative_paths_to_include. use_folders_as_labels (bool, optional): if this is True, class names will be pulled from folder names, useful if you have images like a/b/cat/image001.jpg, a/b/dog/image002.jpg, etc. recursive (bool, optional): whether to recurse into [input_folder] no_json_handling (str, optional): how to deal with image files that have no corresponding .json files, can be: - 'skip': ignore image files with no corresponding .json files - 'empty': treat image files with no corresponding .json files as empty - 'error': throw an error when an image file has no corresponding .json file validate_image_sizes (bool, optional): whether to load images to verify that the sizes specified in the labelme files are correct max_workers (int, optional): number of workers to use for parallelization, set to <=1 to disable parallelization use_threads (bool, optional): whether to use threads (True) or processes (False) for parallelization, not relevant if max_workers <= 1 Returns: dict: a COCO-formatted dictionary, identical to what's written to [output_file] if [output_file] is not None. """ if max_workers > 1: assert category_id_to_category_name is not None, \ 'When parallelizing labelme --> COCO conversion, you must supply a category mapping' if category_id_to_category_name is None: category_name_to_id = {} else: category_name_to_id = {v: k for k, v in category_id_to_category_name.items()} for category_name in category_name_to_id: try: category_name_to_id[category_name] = int(category_name_to_id[category_name]) except ValueError: raise ValueError('Category IDs must be ints or string-formatted ints') # If the user supplied an explicit empty category ID, and the empty category # name is already in category_name_to_id, make sure they match. if empty_category_id is not None: if empty_category_name in category_name_to_id: assert category_name_to_id[empty_category_name] == empty_category_id, \ 'Ambiguous empty category specification' if empty_category_id in category_id_to_category_name: assert category_id_to_category_name[empty_category_id] == empty_category_name, \ 'Ambiguous empty category specification' else: if empty_category_name in category_name_to_id: empty_category_id = category_name_to_id[empty_category_name] del category_id_to_category_name # Enumerate images print('Enumerating images in {}'.format(input_folder)) image_filenames_relative = path_utils.find_images(input_folder, recursive=recursive, return_relative_paths=True, convert_slashes=True) # Remove any images we're supposed to skip if (relative_paths_to_include is not None) or (relative_paths_to_exclude is not None): image_filenames_relative_to_process = [] for image_fn_relative in image_filenames_relative: if relative_paths_to_include is not None and image_fn_relative not in relative_paths_to_include: continue if relative_paths_to_exclude is not None and image_fn_relative in relative_paths_to_exclude: continue image_filenames_relative_to_process.append(image_fn_relative) print('Processing {} of {} images'.format( len(image_filenames_relative_to_process), len(image_filenames_relative))) image_filenames_relative = image_filenames_relative_to_process # If the user supplied a category ID to use for empty images... if empty_category_id is not None: try: empty_category_id = int(empty_category_id) except ValueError: raise ValueError('Category IDs must be ints or string-formatted ints') if empty_category_id is None: empty_category_id = _add_category(empty_category_name,category_name_to_id) print('Processing annotations...') if max_workers <= 1: image_results = [] for image_fn_relative in tqdm(image_filenames_relative): result = _process_labelme_file(image_fn_relative,input_folder,use_folders_as_labels, no_json_handling,validate_image_sizes, category_name_to_id,allow_new_categories=True) image_results.append(result) else: n_workers = min(max_workers,len(image_filenames_relative)) assert category_name_to_id is not None pool = None try: if use_threads: pool = ThreadPool(n_workers) else: pool = Pool(n_workers) image_results = list(tqdm(pool.imap( partial(_process_labelme_file, input_folder=input_folder, use_folders_as_labels=use_folders_as_labels, no_json_handling=no_json_handling, validate_image_sizes=validate_image_sizes, category_name_to_id=category_name_to_id, allow_new_categories=False ),image_filenames_relative), total=len(image_filenames_relative))) finally: if pool is not None: pool.close() pool.join() print('Pool closed and joined for labelme file processing') images = [] annotations = [] # Flatten the lists of images and annotations for result in image_results: im = result['im'] annotations_this_image = result['annotations_this_image'] if im is None: assert annotations_this_image is None else: images.append(im) annotations.extend(annotations_this_image) output_dict = {} output_dict['images'] = images output_dict['annotations'] = annotations if info_struct is None: info_struct = {} if 'description' not in info_struct: info_struct['description'] = \ 'Converted to COCO from labelme annotations in folder {}'.format(input_folder) if 'version' not in info_struct: info_struct['version'] = 1.0 output_dict['info'] = info_struct categories = [] for category_name in category_name_to_id: categories.append({'name':category_name,'id':category_name_to_id[category_name]}) output_dict['categories'] = categories if output_file is not None: with open(output_file,'w') as f: json.dump(output_dict,f,indent=1) return output_dict
# ...def labelme_to_coco()
[docs] def find_empty_labelme_files(input_folder,recursive=True): """ Returns a list of all image files in in [input_folder] associated with .json files that have no boxes in them. Also returns a list of images with no associated .json files. Specifically, returns a dict: .. code-block: none { 'images_with_empty_json_files':[list], 'images_with_no_json_files':[list], 'images_with_non_empty_json_files':[list] } Args: input_folder (str): the folder to search for empty (i.e., box-less) Labelme .json files recursive (bool, optional): whether to recurse into [input_folder] Returns: dict: a dict with fields: - images_with_empty_json_files: a list of all image files in [input_folder] associated with .json files that have no boxes in them - images_with_no_json_files: a list of images in [input_folder] with no associated .json files - images_with_non_empty_json_files: a list of images in [input_folder] associated with .json files that have at least one box """ image_filenames_relative = path_utils.find_images(input_folder, recursive=recursive, return_relative_paths=True) images_with_empty_json_files = [] images_with_no_json_files = [] images_with_non_empty_json_files = [] # fn_relative = image_filenames_relative[0] for fn_relative in image_filenames_relative: image_fn_abs = os.path.join(input_folder,fn_relative) json_fn_abs = os.path.splitext(image_fn_abs)[0] + '.json' if not os.path.isfile(json_fn_abs): images_with_no_json_files.append(fn_relative) continue else: # Read the .json file with open(json_fn_abs,'r') as f: labelme_data = json.load(f) shapes = labelme_data['shapes'] if len(shapes) == 0: images_with_empty_json_files.append(fn_relative) else: images_with_non_empty_json_files.append(fn_relative) # ...for every image return {'images_with_empty_json_files':images_with_empty_json_files, 'images_with_no_json_files':images_with_no_json_files, 'images_with_non_empty_json_files':images_with_non_empty_json_files}
# ...def find_empty_labelme_files(...) #%% Interactive driver if False: pass #%% Options empty_category_name = 'empty' empty_category_id = None category_id_to_category_name = None info_struct = None input_folder = os.path.expanduser('~/data/md-test') output_file = os.path.expanduser('~/data/md-test-labelme-to-coco.json') #%% Programmatic execution output_dict = labelme_to_coco(input_folder,output_file, category_id_to_category_name=category_id_to_category_name, empty_category_name=empty_category_name, empty_category_id=empty_category_id, info_struct=None, use_folders_as_labels=False, validate_image_sizes=False, no_json_handling='empty') #%% Validate from megadetector.data_management.databases import integrity_check_json_db options = integrity_check_json_db.IntegrityCheckOptions() options.baseDir = input_folder options.bCheckImageSizes = True options.bCheckImageExistence = True options.bFindUnusedImages = True options.bRequireLocation = False sorted_categories, _, error_info = integrity_check_json_db.integrity_check_json_db(output_file,options) #%% Preview from megadetector.visualization import visualize_db options = visualize_db.DbVizOptions() options.parallelize_rendering = True options.viz_size = (900, -1) options.num_to_visualize = 5000 html_file,_ = visualize_db.visualize_db(output_file,os.path.expanduser('~/tmp/labelme_to_coco_preview'), input_folder,options) from megadetector.utils import path_utils # noqa path_utils.open_file(html_file) #%% Prepare command line s = 'python labelme_to_coco.py {} {}'.format(input_folder,output_file) print(s) import clipboard; clipboard.copy(s) #%% Command-line driver def main(): # noqa parser = argparse.ArgumentParser( description='Convert labelme-formatted data to COCO') parser.add_argument( 'input_folder', type=str, help='Path to images and .json annotation files') parser.add_argument( 'output_file', type=str, help='Output filename (.json)') if len(sys.argv[1:]) == 0: parser.print_help() parser.exit() args = parser.parse_args() labelme_to_coco(args.input_folder,args.output_file) if __name__ == '__main__': main()