Labelbox to YOLOnet

labelbox_label.py

import sys, os
import json
import glob
from PIL import Image

USAGE = "python labelbox_label.py /path/to/training/image/directory/ /path/to/labelbox/exported/xy/data.json"

assert len(sys.argv) > 2, f'Insufficient arguments: {USAGE}'
args = sys.argv[1:] if '.py' in sys.argv[0] else sys.argv[0:]
assert(len(args) == 2), f'Incorrect usage: {USAGE}'

IMAGE_DIR = args[0]
JSON_FILE = args[1]

with open(JSON_FILE) as f:
    data = json.load(f)
    print(f'Found {len(data)} datapoints.')
images = glob.glob(f'{IMAGE_DIR}/*.jpg')
print(f'Found {len(images)} images.')

classes = list(set(str(key) for d in data for key in d["Label"].keys()))
classes.remove('START')
print(f'Found {len(classes)} classes.')

def convert_bbox(bbox, width, height, offset_x=0, offset_y=0):
    x = min([coord["x"] for coord in bbox])
    y = min([coord["y"] for coord in bbox])
    w = max([coord["x"] for coord in bbox]) - x
    h = max([coord["y"] for coord in bbox]) - y
    x += w/2
    y += h/2
    x -= max(offset_x, 0)
    y = max(offset_y - y, 0)
    return x/width, y/height, w/width, h/height

for image in images:
    # Trim the extension for our use case
    image = image.split('/')[-1]
    # Get the width and height of the original image
    img = Image.open(f'{IMAGE_DIR}/{image}')
    width, height = img.size[:2]
    width, height = int(width), int(height)
    # Next we look up the image in the JSON data
    annotation = [d for d in data if image.split(' ')[0] in d["Labeled Data"]][0]
    with open(f'{IMAGE_DIR}/{image[:image.rfind(".")]}.txt', 'w') as label:
        startbox = annotation["Label"]["START"][0]
        offset_x, offset_y = startbox["x"], startbox["y"]
        for bbox in annotation["Label"].keys():
            if bbox == "START": continue
            x, y, w, h = convert_bbox(annotation["Label"][bbox][0], width, height, offset_x, offset_y)
            label.write(f'{classes.index(bbox)} {x} {y} {w} {h}\n')
    print(f'{image}: {len(annotation["Label"].keys())} annotations')

with open(f'{IMAGE_DIR}/train.txt', 'w') as image_paths:
    for image in images:
        image_paths.write(f'{image}\n')

with open(f'{IMAGE_DIR}/custom.names', 'w') as names:
    for classname in classes:
        names.write(f'{classname}\n')

with open(f'{IMAGE_DIR}/custom.cfg', 'w') as config:
    config.write(f'classes={len(classes)}\n')
    config.write(f'train={IMAGE_DIR}/train.txt\n')
    config.write(f'valid={IMAGE_DIR}/train.txt\n') # TODO: add support for validation set
    config.write(f'names={IMAGE_DIR}/custom.names\n')
    config.write(f'backup=backup') # TODO: add custom checkpointing directory

print("Change the classes at the bottom of and in the middle of the architecture configuration you're using to match the number of classes you have.")
print("In the same file also change filters of the first conv. layer, the value should be num/3*(classes+5)")

I unfortunately wrote this script a very long time ago so Labelbox may have changed their output in the meantime. If there is no longer a start token in the output you can just remove this line :)