# -*- coding: utf-8 -*- # @Author : LG from json import load, dump import os from pycocotools import mask as coco_mask import cv2 import imgviz class COCOConverter: def __init__(self, ): pass # labels = cfg.get('label', []) # for index, label_dict in enumerate(labels): # category = label_dict.get('name', 'unknow') # color = label_dict.get('color', '#000000') def convert_to_coco(self, isat_json_root:str, to_path:str): coco_anno = {} # info coco_anno['info'] = {} coco_anno['info']['description'] = 'coco from ISAT' coco_anno['info']['version'] = None coco_anno['info']['year'] = None coco_anno['info']['contributor'] = None coco_anno['info']['date_created'] = None # licenses coco_anno['licenses'] = [] coco_anno['licenses'][0] = {} coco_anno['licenses'][0]['url'] = None coco_anno['licenses'][0]['id'] = 0 coco_anno['licenses'][0]['name'] = None # images and annotations coco_anno['images'] = [] coco_anno['annotations'] = [] jsons = [f for f in os.listdir(isat_json_root) if f.endswith('.json')] for json in jsons: coco_image_info = {} dataset = load(os.path.join(isat_json_root, json)) info = dataset.get('info', {}) description = info.get('description', '') if not description.startswith('ISAT'): # 不是ISAT格式json continue img_name = info.get('name', '') width = info.get('width', None) height = info.get('height', None) depth = info.get('depth', None) note = info.get('note', '') objects = dataset.get('objects', []) for obj in objects: category = obj.get('category', 'unknow') group = obj.get('group', 0) if group is None: group = 0 segmentation = obj.get('segmentation', []) iscrowd = obj.get('iscrowd', 0) note = obj.get('note', '') area = obj.get('area', 0) layer = obj.get('layer', 2) bbox = obj.get('bbox', []) def convert_from_coco(self, coco_json:str, to_path:str, keep_crowd:bool=False): assert coco_json.endswith('.json') annos = {} if os.path.exists(coco_json): with open(coco_json, 'r') as f: dataset = load(f) images = {image.get('id', None):{ 'file_name': image.get('file_name', ''), 'height': image.get('height', ''), 'width': image.get('width', ''), } for image in dataset.get('images', [])} annotations = dataset.get('annotations', []) categories = {categorie.get('id', None): {'name': categorie.get('name', '')} for categorie in dataset.get('categories', [])} for index, annotation in enumerate(annotations): annotation_index = annotation.get('id') annotation_image_id = annotation.get('image_id') annotation_category_id = annotation.get('category_id') file_name = images[annotation_image_id].get('file_name') height = images[annotation_image_id].get('height') width = images[annotation_image_id].get('width') iscrowd = annotation["iscrowd"] if file_name == '000000279278.jpg': continue if annotation_image_id not in annos: annos[annotation_image_id] = {} objects = annos[annotation_image_id].get('objects', []) if iscrowd == 0: # polygon segmentations = annotation.get('segmentation') for segmentation in segmentations: xs = segmentation[::2] ys = segmentation[1::2] points = [[x, y] for x ,y in zip(xs, ys)] obj = { 'category': categories.get(annotation_category_id).get('name'), 'group': annotation_index, 'area': None, 'segmentation': points, 'layer': 1, 'bbox': None, 'iscrowd': iscrowd, 'note': '' } objects.append(obj) elif iscrowd == 1 and keep_crowd: # RLE segmentations = annotation.get('segmentation', {}) rles = coco_mask.frPyObjects(segmentations, height, width) masks = coco_mask.decode(rles) contours, _ = cv2.findContours(masks, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_TC89_KCOS) for contour in contours: points = [] for point in contour: x, y = point[0] points.append([float(x), float(y)]) obj = { 'category': categories.get(annotation_category_id).get('name'), 'group': annotation_index, 'area': None, 'segmentation': points, 'layer': 1, 'bbox': None, 'iscrowd': iscrowd, 'note': '' } objects.append(obj) else: pass annos[annotation_image_id]['objects'] = objects for image_id, values in annos.items(): image_path = images[image_id].get('file_name') folder, name = os.path.split(image_path) height = images[image_id].get('height') width = images[image_id].get('width') objects = values.get('objects', []) isat_anno = {} isat_anno['info'] = {} isat_anno['info']['description'] = 'ISAT' isat_anno['info']['folder'] = folder isat_anno['info']['name'] = name isat_anno['info']['width'] = width isat_anno['info']['height'] = height isat_anno['info']['depth'] = None isat_anno['info']['note'] = '' isat_anno['objects'] = [] # coco annotation的id 太大了，这里缩一下，每张图片重新开始计数 groups_dict = {} for obj in objects: group = obj.get('group', 0) if group not in groups_dict: groups_dict[group] = len(groups_dict)+1 for obj in objects: object = {} object['category'] = obj.get('category', '') object['group'] = groups_dict.get(obj.get('group', 0)) object['segmentation'] = obj.get('segmentation', []) object['area'] = obj.get('area', None) object['layer'] = obj.get('layer', None) object['bbox'] = obj.get('bbox', None) object['iscrowd'] = obj.get('iscrowd', 0) object['note'] = obj.get('note', '') isat_anno['objects'].append(object) json_name = '.'.join(name.split('.')[:-1]) + '.json' save_json = os.path.join(to_path, json_name) with open(save_json, 'w') as f: try: dump(isat_anno, f) print('Converted coco to ISAT: {}'.format(json_name)) except Exception as e: print('Convert coco to ISAT {} ,error: {}'.format(json_name, e)) ### 类别文件 cmap = imgviz.label_colormap() if __name__ == '__main__': coco = COCOConverter() coco.convert_from_coco( '/mnt/disk/coco/annotations/instances_val2017.json', '/mnt/disk/coco/isat_json', True )