"""Generate BONAI official_v1 manifests and COCO subsets from released annotations.""" import argparse from collections import Counter, defaultdict from copy import deepcopy import hashlib import json from pathlib import Path import random import re CITIES = ('beijing', 'chengdu', 'haerbin', 'jinan', 'shanghai') SEED = 42 VERSION = 'official_v1' TEST_FILE = 'coco/bonai_shanghai_xian_test.json' SOURCE_SHA256 = { 'coco/bonai_beijing_trainval.json': 'f495f2d3a7aecbca0466b2f54ee03d0a885e8c6da16e19d56ce74d5bd7ccfaa7', 'coco/bonai_chengdu_trainval.json': '30d2f0848101d602a3e2225631a4e93cc2612e96057e2b70ba0f22416bb2a7bb', 'coco/bonai_haerbin_trainval.json': '2e95d0251060ab7750c2ad4f14d7c6e9d55ba91fbed7217d945b4baea589c445', 'coco/bonai_jinan_trainval.json': '7b0ce3cc30603b9efced9603188ae5b01e93304c78da82c9f6d826c994098c98', 'coco/bonai_shanghai_trainval.json': '78f609ced48f2886ba3683dfdc5c5643354d26b50df4e26c7611b1c7b04551cc', 'coco/bonai_shanghai_xian_test.json': '8b492c2c685758f1f697a9883e31a9e913d0a02926cdd1a57f9527f6b371fe01', } def tile(filename): match = re.search(r'L(\d+)_(\d+)_(\d+)__(\d+)_(\d+)\.png$', filename) if match is None: raise ValueError(f'Unsupported image filename: {filename}') return tuple(map(int, match.groups()[:3])) def dump(path, data): path.parent.mkdir(parents=True, exist_ok=True) path.write_text(json.dumps(data, ensure_ascii=False, indent=2) + '\n') def generate(data_root, output_root): sources, hashes = {}, {} for city in CITIES: name = f'coco/bonai_{city}_trainval.json' content = (data_root / name).read_bytes() sources[city] = json.loads(content) hashes[name] = hashlib.sha256(content).hexdigest() test_bytes = (data_root / TEST_FILE).read_bytes() test = json.loads(test_bytes) hashes[TEST_FILE] = hashlib.sha256(test_bytes).hexdigest() if hashes != SOURCE_SHA256: raise ValueError('Source files differ from the official_v1 inputs; see SOURCE_SHA256.') groups = defaultdict(list) for image in sources['shanghai']['images']: groups[tile(image['file_name'])].append(image) assert len(groups) == 157 expected = {(source, crop) for source in ('arg', 'google', 'ms') for crop in ('0_0.png', '0_1024.png', '1024_0.png', '1024_1024.png')} for images in groups.values(): assert len(images) == 12 assert {(im['file_name'].split('__')[0].split('_')[-1], im['file_name'].split('__')[-1]) for im in images} == expected validation_tiles = set(random.Random(SEED).sample(sorted(groups), 19)) template = {k: deepcopy(v) for k, v in sources['shanghai'].items() if k not in ('images', 'annotations')} outputs = {} for split in ('train', 'val'): outputs[split] = deepcopy(template) outputs[split]['info']['description'] = f'BONAI {VERSION} {split}' outputs[split]['images'] = [] outputs[split]['annotations'] = [] mapping, annotation_ranges = [], {} next_image_id, next_annotation_id = 1, 1 for city in CITIES: source = sources[city] assert source['categories'] == template['categories'] assert source['licenses'] == template['licenses'] source_name = f'coco/bonai_{city}_trainval.json' remap, membership = {}, {} for image in sorted(source['images'], key=lambda im: im['id']): split = 'val' if city == 'jinan' or ( city == 'shanghai' and tile(image['file_name']) in validation_tiles) else 'train' assert image['id'] not in remap remap[image['id']] = next_image_id membership[image['id']] = split outputs[split]['images'].append({**image, 'id': next_image_id}) mapping.append(dict(file_name=image['file_name'], source_file=source_name, source_image_id=image['id'], image_id=next_image_id, split=split)) next_image_id += 1 annotation_ranges[source_name] = dict(first_id=next_annotation_id, count=len(source['annotations'])) seen_annotation_ids = set() for annotation in sorted(source['annotations'], key=lambda ann: ann['id']): assert annotation['id'] not in seen_annotation_ids seen_annotation_ids.add(annotation['id']) old_image_id = annotation['image_id'] split = membership[old_image_id] outputs[split]['annotations'].append( {**annotation, 'id': next_annotation_id, 'image_id': remap[old_image_id]}) next_annotation_id += 1 names = {split: {im['file_name'] for im in data['images']} for split, data in outputs.items()} names['test'] = {im['file_name'] for im in test['images']} all_source_names = {im['file_name'] for source in sources.values() for im in source['images']} assert len(all_source_names) == 3000 assert len(names['train']) == 2700 and len(names['val']) == 300 and len(names['test']) == 300 assert names['train'] | names['val'] == all_source_names split_groups = {split: {tile(name) for name in files} for split, files in names.items()} for a, b in [('train', 'val'), ('train', 'test'), ('val', 'test')]: assert not names[a] & names[b] assert not split_groups[a] & split_groups[b] # Parent tiles cover 2048 pixels = 8 tiles on the 256-pixel L18 grid. assert not any(x[0] == y[0] and abs(x[1] - y[1]) < 8 and abs(x[2] - y[2]) < 8 for x in split_groups[a] for y in split_groups[b]) folder = output_root / 'splits' / VERSION folder.mkdir(parents=True, exist_ok=True) for split, files in names.items(): (folder / f'{split}.txt').write_text(''.join(name + '\n' for name in sorted(files))) for split, data in outputs.items(): path = output_root / 'coco' / f'bonai_{split}_{VERSION}.json' path.parent.mkdir(parents=True, exist_ok=True) path.write_text(json.dumps(data, ensure_ascii=False, separators=(',', ':')) + '\n') dump(folder / 'image_id_map.json', mapping) protocol = dict( version=VERSION, seed=SEED, selection='random.Random(42).sample(sorted(Shanghai parent tile tuples), 19)', grouping='(zoom, first tile coordinate, second tile coordinate); all sources and crops stay together', shanghai_validation_tiles=['L%d_%d_%d' % group for group in sorted(validation_tiles)], source_sha256=hashes, counts={split: dict(images=len(data['images']), annotations=len(data['annotations']), cities=dict(Counter(im['file_name'].split('_')[0] if im['file_name'].startswith(tuple(CITIES)) else 'xian' for im in data['images']))) for split, data in {**outputs, 'test': test}.items()}, annotation_id_ranges=annotation_ranges, id_policy='Images and annotations are enumerated from 1 across alphabetically ordered source cities, then original numeric ID; train and val retain these global IDs. Test IDs are unchanged.', geographic_check='No shared parent tile IDs or positive-area overlap between filename-derived 8-by-8 parent tile boxes across splits. Adjacent tile boundaries are allowed; no geographic buffer is imposed. This does not independently verify imagery georeferencing.', provenance='This manifest defines the official_v1 protocol. Its identity with the historical image-level partition used for the paper results has not been established.', ) dump(folder / 'protocol.json', protocol) print(json.dumps(protocol['counts'], indent=2)) if __name__ == '__main__': parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('--data-root', type=Path, required=True, help='Dataset root containing the original coco/ files') parser.add_argument('--output-root', type=Path, required=True) args = parser.parse_args() generate(args.data_root, args.output_root)