utils/convert_masks.py

#!/bin/env python

# Mask conversion for plantorgans
# 2024 by Jiri Podivin

import numpy as np
import json
import os
import numpy as np
from PIL import Image
from multiprocessing import Pool, cpu_count
from urllib.parse import unquote
from datetime import datetime
import pandas as pd
import os
import tempfile
import argparse
import glob
import shutil
import tarfile

class InputStream:
    def __init__(self, data):
        self.data = data
        self.i = 0

    def read(self, size):
        out = self.data[self.i : self.i + size]
        self.i += size
        return int(out, 2)


def access_bit(data, num):
    """from bytes array to bits by num position"""
    base = int(num // 8)
    shift = 7 - int(num % 8)
    return (data[base] & (1 << shift)) >> shift


def bytes2bit(data):
    """get bit string from bytes data"""
    return ''.join([str(access_bit(data, i)) for i in range(len(data) * 8)])


def decode_rle(rle, print_params: bool = False):
    """from LS RLE to numpy uint8 3d image [width, height, channel]

    Args:
        print_params (bool, optional): If true, a RLE parameters print statement is suppressed
    """
    input = InputStream(bytes2bit(rle))
    num = input.read(32)
    word_size = input.read(5) + 1
    rle_sizes = [input.read(4) + 1 for _ in range(4)]

    if print_params:
        print(
            'RLE params:', num, 'values', word_size, 'word_size', rle_sizes, 'rle_sizes'
        )

    i = 0
    out = np.zeros(num, dtype=np.uint8)
    while i < num:
        x = input.read(1)
        j = i + 1 + input.read(rle_sizes[input.read(2)])
        if x:
            val = input.read(word_size)
            out[i:j] = val
            i = j
        else:
            while i < j:
                val = input.read(word_size)
                out[i] = val
                i += 1
    return out

def log(message):
    timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
    print(f"[{timestamp}] {message}")

def save_image(mask_image: Image.Image, save_path: str):
    mask_image.save(save_path, format='PNG')
    log(f'Saved mask: {save_path}')

def process_files_in_parallel(files_to_process, masks_save_directory, source_files):
    with Pool(processes=cpu_count()//2) as pool:
        results = pool.starmap(process_file, [(file, masks_save_directory, source_files) for file in files_to_process])
    
    return [e for r in results for e in r]


def process_file(file_path, masks_save_directory, source_files):
    log(f"Opening file: {file_path}")
    total_metadata = []
    try:
        with open(file_path, 'r') as file:
            data = json.load(file)
    except Exception as e:
        log(f'Error reading file {file_path}: {e}')
        return total_metadata
    
    image_name = data['task']['data']['image'].split('/')[-1]

    if image_name not in source_files:
        log(f"Requested file {image_name} does not exist in source data!")
        return total_metadata

    image_name_prefix = unquote(image_name.rsplit('.', 1)[0])
    log(f"Processing image: {image_name_prefix}")
    label_counts = {}

    for result in data['result']:
        if 'rle' not in result['value']:
            log(f"No 'rle' key found in result: {result.get('id', 'Unknown ID')}")
            continue
        
        rle_data = result['value']['rle']
        rle_bytes = bytes.fromhex(''.join(format(x, '02x') for x in rle_data))
        mask = decode_rle(rle_bytes)

        original_height = result['original_height']
        original_width = result['original_width']
        mask = mask.reshape((original_height, original_width, 4))

        alpha_channel = mask[:, :, 3]
        mask_image = np.zeros((original_height, original_width, 3), dtype=np.uint8)
        mask_image[alpha_channel == 255] = [255, 255, 255]

        if 'brushlabels' in result['value']:
            for label in result['value']['brushlabels']:

                label_counts[label] = label_counts.get(label, 0) + 1
                save_path = os.path.join(masks_save_directory, f"{image_name_prefix}-{label}-{label_counts[label]}.png")
                save_image(Image.fromarray(mask_image).convert('L'), save_path)
                metadata = {
                    "original_height": result['original_height'],
                    "original_width": result['original_width'],
                    "image": os.path.join('sourcedata/labeled/', os.path.basename(data['task']['data']['image'])),
                    "score": result['score'] if 'score' in result.keys() else 0,
                    "mask": save_path,
                    "class": label,
                }
                total_metadata.append(metadata)

    return total_metadata

def merge_file_masks(mask_info, target_mask_dir, label2id, img):
    final_mask = np.zeros(
        np.asarray(Image.open(mask_info['mask'].iloc[0])).shape, dtype=np.uint8)
    for i, r in mask_info.iterrows():
        mask = np.asarray(Image.open(r['mask']))
        final_mask = np.where(mask == 0, final_mask, label2id[r['class']])
    
    mask_path = os.path.join(target_mask_dir, f"{os.path.basename(img).split('.')[0]}_mask.png")
    Image.fromarray(final_mask).convert('L').save(mask_path, format='PNG')
    return {
        'mask': mask_path,
        'image': img,
        'original_height': r['original_height'],
        'original_width': r['original_width']
    }

def merge_masks(mask_metadata, target_mask_dir, label2id):
    new_metadata = []
    imgs = [
        (
            mask_metadata[mask_metadata['image'] == img],
            target_mask_dir,
            label2id,
            img
        ) for img in mask_metadata['image'].unique()]

    
    with Pool(processes=cpu_count()//2) as pool:
        new_metadata = pool.starmap(merge_file_masks, imgs)

    return new_metadata


def main():
    parser = argparse.ArgumentParser('maskconvert')
    parser.add_argument('dataset_root', help="Root directory of the dataset repo.")

    arguments = parser.parse_args()

    # Unzip raw labels to temporary location
    tmp_raw_label_pth = tempfile.mkdtemp('_labels_raw')
    with tarfile.open(os.path.join(arguments.dataset_root, 'labels_raw.tar.gz'), mode='r:gz') as archive:
        archive.extractall(tmp_raw_label_pth)

    annotations_folder_path = os.path.join(tmp_raw_label_pth, 'labels_raw')
    
    tmp_mask_path = tempfile.mkdtemp('_masks')
 
    files_to_process = glob.glob(f"{annotations_folder_path}/*")

    # Image name -> path dict
    images = {os.path.basename(x): x for x in glob.glob('sourcedata/**/*.jpg')}

    # Image names
    source_files = [k for k in images.keys()]

    metadata = pd.DataFrame(process_files_in_parallel(files_to_process, tmp_mask_path, source_files))

    # Label map order IS important
    id2label = {int(k): v for k, v in enumerate(['void', 'Fruit', 'Leaf', 'Flower', 'Stem'])}

    label2id = {v: k for k, v in id2label.items()}

    result = merge_masks(metadata, os.path.join(arguments.dataset_root, 'semantic_masks'), label2id)

    result = pd.DataFrame(result).drop_duplicates()
    
    # Moving newly labeled images to right dir
    images = {os.path.basename(x): x for x in glob.glob('sourcedata/**/*.jpg')}

    for img in metadata['image']:
        if 'unlabeled' in images[img]:
            print("Moving {img} to labeled!")
            shutil.move(images[img], os.path.join('sourcedata/labeled/', img))

    result.to_csv(
        os.path.join(arguments.dataset_root, 'semantic_metadata.csv'),
        index=False)


if __name__ == '__main__':
    main()