将分割数据中的像素级多边形标注转换为目标检测的边界框（bbox）标注，其实并不复杂。每个分割目标的多边形标注由多个像素点的 [x, y] 坐标组成，只需找到这些坐标的 Xmin, Ymin, Xmax, 和 Ymax 值，就能确定该目标的边界框。

结果：

preprocess.py代码如下

import os
import glob
from shutil import copy2
from PIL import Image
import json
import numpy as npdef copy_file(src, src_ext, dst):# find all files ends up with extflist = sorted(glob.glob(os.path.join(src, '*', src_ext)))for fname in flist:src_path = os.path.join(src, fname)copy2(src_path, dst)print('copied %s to %s' % (src_path, dst))def construct_box(inst_root, inst_name, cls_name, dst):inst_list = sorted(glob.glob(os.path.join(inst_root, '*', inst_name)))cls_list = sorted(glob.glob(os.path.join(inst_root, '*', cls_name)))for inst, cls in zip(*(inst_list, cls_list)):inst_map = Image.open(os.path.join(inst_root, inst))inst_map = np.array(inst_map, dtype=np.int32)cls_map = Image.open(os.path.join(inst_root, cls))cls_map = np.array(cls_map, dtype=np.int32)H, W = inst_map.shape# get a list of unique instancesinst_info = {'imgHeight':H, 'imgWidth':W, 'objects':{}}inst_ids = np.unique(inst_map)for iid in inst_ids: if int(iid) < 1000: # filter out non-instance maskscontinueys,xs = np.where(inst_map==iid)ymin, ymax, xmin, xmax = \ys.min(), ys.max(), xs.min(), xs.max()cls_label = np.median(cls_map[inst_map==iid])inst_info['objects'][str(iid)] = {'bbox': [int(xmin), int(ymin), int(xmax), int(ymax)],'cls': int(cls_label)}# write a file to pathfilename = os.path.splitext(os.path.basename(inst))[0]savename = os.path.join(dst, filename + '.json')with open(savename, 'w') as f:json.dump(inst_info, f)print('wrote a bbox summary of %s to %s' % (inst, savename))# organize image
if __name__ == '__main__':folder_name = 'datasets/cityscape/'train_img_dst = os.path.join(folder_name, 'train_img')train_label_dst = os.path.join(folder_name, 'train_label')train_inst_dst = os.path.join(folder_name, 'train_inst')train_bbox_dst = os.path.join(folder_name, 'train_bbox')val_img_dst = os.path.join(folder_name, 'val_img')val_label_dst = os.path.join(folder_name, 'val_label')val_inst_dst = os.path.join(folder_name, 'val_inst')val_bbox_dst = os.path.join(folder_name, 'val_bbox')if not os.path.exists(train_img_dst):os.makedirs(train_img_dst)if not os.path.exists(train_label_dst):os.makedirs(train_label_dst)if not os.path.exists(train_inst_dst):os.makedirs(train_inst_dst)if not os.path.exists(val_img_dst):os.makedirs(val_img_dst)if not os.path.exists(val_label_dst):os.makedirs(val_label_dst)if not os.path.exists(val_inst_dst):os.makedirs(val_inst_dst)root = "G:\Dataset\cityscapes"# # train_image# copy_file('G:/Dataset/cityscapes/leftImg8bit/train',\#         '*_leftImg8bit.png', train_img_dst)# # train_label# copy_file('G:/Dataset/cityscapes/gtFine/train',\#         '*_labelIds.png', train_label_dst)# # train_inst# copy_file('G:/Dataset/cityscapes/gtFine/train',\#         '*_instanceIds.png', train_inst_dst)# # val_image# copy_file('G:/Dataset/cityscapes/leftImg8bit/val',\#         '*_leftImg8bit.png', val_img_dst)# # val_label# copy_file('G:/Dataset/cityscapes/gtFine/val',\#         '*_labelIds.png', val_label_dst)# # val_inst# copy_file('G:/Dataset/cityscapes/gtFine/val',\#         '*_instanceIds.png', val_inst_dst)if not os.path.exists(train_bbox_dst):os.makedirs(train_bbox_dst)if not os.path.exists(val_bbox_dst):os.makedirs(val_bbox_dst)# wrote a bounding box summary construct_box('G:/Dataset/cityscapes/gtFine/train',\'*_instanceIds.png', '*_labelIds.png', train_bbox_dst)construct_box('G:/Dataset/cityscapes/gtFine/val',\'*_instanceIds.png', '*_labelIds.png', val_bbox_dst)