diff --git a/prepare_data/python/prepare_data.py b/prepare_data/python/prepare_data.py new file mode 100644 index 0000000..c70a073 --- /dev/null +++ b/prepare_data/python/prepare_data.py @@ -0,0 +1,164 @@ +# -* coding: utf-8 -*- +""" +------------------------------------------------- + Description : An python example of preparing data for da-faster-rcnn, the exmample is based on adaptation from Cityscapes + to Foggy Cityscapes, other datasets can be prepared similarly. + Author : kinredon + date： 2019/2/26 +------------------------------------------------- +""" +import os +import cv2 +import numpy as np +import shutil +from writexml import writexml +from tqdm import tqdm + +# specify path +source_data_dir = 'CITYSCAPES_DIR' +target_data_dir = 'FOGGY_CITYSCAPES_DIR' + +# initialization +img_dir = 'VOCdevkit2007/VOC2007/JPEGImages' +sets_dir = 'VOCdevkit2007/VOC2007/ImageSets/Main' +annotation_dir = 'VOCdevkit2007/VOC2007/Annotations' + +if not os.path.exists(img_dir): + os.makedirs(img_dir) + +if not os.path.exists(sets_dir): + os.makedirs(sets_dir) + +if not os.path.exists(annotation_dir): + os.makedirs(annotation_dir) + +# organize images & prepare split list. + +# process source train images +source_img_path = os.path.join(source_data_dir, 'leftImg8bit', 'train') +source_train_list = [] +print('process source train images') +for root, dirs, files in os.walk(source_img_path, topdown=True): + files = [name for name in files if name.endswith('.png')] + for name in tqdm(files): + im_name = name.replace('.png', '') + im_name = 'source_' + im_name + img = cv2.imread(os.path.join(root, name)) + cv2.imwrite(os.path.join(img_dir, im_name + '.jpg'), img) + source_train_list.append(im_name) + +# process target train images. +print('process target train images') +target_img_path = os.path.join(target_data_dir, 'leftImg8bit_foggy', 'train') +target_train_list = [] +for root, dirs, files in os.walk(target_img_path, topdown=True): + files = [ name for name in files if name.endswith('_foggy_beta_0.02.png')] + for name in tqdm(files): + + im_name = name.replace('.png', '') + im_name = 'target_' + im_name + + img = cv2.imread(os.path.join(root, name)) + cv2.imwrite(os.path.join(img_dir, im_name + '.jpg'), img) + target_train_list.append(im_name) + + +# process target test images +print('process target test images') +target_img_path = os.path.join(target_data_dir, 'leftImg8bit_foggy', 'val') +target_test_list = [] +for root, dirs, files in os.walk(target_img_path, topdown=True): + files = [name for name in files if name.endswith('_foggy_beta_0.02.png')] + for name in tqdm(files): + + im_name = name.replace('.png', '') + im_name = 'target_' + im_name + + img = cv2.imread(os.path.join(root, name)) + cv2.imwrite(os.path.join(img_dir, im_name + '.jpg'), img) + target_test_list.append(im_name) + +# write the list +train_list = source_train_list + target_train_list +test_list = target_test_list + +with open(os.path.join(sets_dir, "trainval.txt"), 'w') as f: + for item in train_list: + f.write("{}\n".format(item)) + +with open(os.path.join(sets_dir, "test.txt"), 'w') as f: + for item in test_list: + f.write("{}\n".format(item)) + + +# prepare the annotation needed for training/testing. +cityscapes_semantics = ['ego vehicle', 'rectification border', 'out of roi', 'static', 'dynamic', 'ground', 'road', + 'sidewalk', 'parking', 'rail track', 'building', 'wall', 'fence', 'guard rail', 'bridge', + 'tunnel', 'pole', 'polegroup', 'traffic light', 'traffic sign', 'vegetation', 'terrain', 'sky', + 'person', 'rider', 'car', 'truck', 'bus', 'caravan', 'trailer', 'train', 'motorcycle', 'bicycle'] +instance_semantics = ['person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle', 'bicycle'] + +bind = {} +for i, elt in enumerate(instance_semantics): + if elt not in bind: + bind[elt] = i +lb_filter = [bind.get(itm, -1) for itm in cityscapes_semantics] + + + +# instanceIds.png +source_img_path = os.path.join(source_data_dir, 'gtFine') +for root, dirs, files in os.walk(source_img_path, topdown=True): + files = [name for name in files if name.endswith('instanceIds.png')] + for name in tqdm(files): + + im_name = name.replace('_gtFine_instanceIds.png', '_leftImg8bit') + im_name = 'source_' + im_name + + im_inst = cv2.imread(os.path.join(root, name), cv2.IMREAD_ANYDEPTH) + im_lb = cv2.imread(os.path.join(root, name.replace('_gtFine_instanceIds.png', '_gtFine_labelIds.png')), cv2.IMREAD_ANYDEPTH) + + all_inst_id = np.setdiff1d(np.unique(im_inst), 0) + boxes = [] + categories = [] + for i_inst in all_inst_id: + inst_mask = (im_inst == i_inst) + inst_mask_int = inst_mask - 0 + # print(np.unique(inst_mask)) + assert (len(np.unique(inst_mask_int)) == 2) + + x_cods = np.where(np.sum(inst_mask_int, 0) > 0) + y_cods = np.where(np.sum(inst_mask_int, 1) > 0) + box = [np.min(x_cods), np.min(y_cods), np.max(x_cods), np.max(y_cods)] + boxes.append(box) + # print(box) + inst_lb = np.unique(im_lb[inst_mask]) + # print(inst_lb) + category = lb_filter[inst_lb[0] - 1] + categories.append(category + 1) + # break + + # plus 1, in order to match matlab code + boxes = np.array(boxes) + 1 + categories = np.array(categories) + boxes = boxes[categories != 0] + categories = categories[categories != 0] + filename = im_name + '.jpg' + img_shape = (*im_inst.shape, 3) + writexml(filename, img_shape, boxes, categories, instance_semantics, + os.path.join(annotation_dir, "{}.xml".format(im_name))) + +all_target_images = target_train_list + target_test_list +for im_name in all_target_images: + source_name = (im_name.replace('_foggy_beta_0.02', '')).replace('target', 'source') + shutil.copyfile(os.path.join(annotation_dir, "{}.xml".format(source_name)), + os.path.join(annotation_dir, "{}.xml".format(im_name))) +print("done!") +# you need move generated 'VOCdevkit2007' directory into $FRCN_ROOT/data + + + + + + + diff --git a/prepare_data/python/writexml.py b/prepare_data/python/writexml.py new file mode 100644 index 0000000..728b492 --- /dev/null +++ b/prepare_data/python/writexml.py @@ -0,0 +1,101 @@ +# -*- coding: utf-8 -*- +""" +------------------------------------------------- + Description : annotations write to xml file + Author : kinredon + date： 2019/2/27 +------------------------------------------------- +""" + +from xml.dom.minidom import Document + + +def writexml(filename, img_shape, bboxes, categories, instance_semantics, xmlpath): + """ + + write to xml style of VOC dataset + :param filename: xml filename + :param img_shape: the image shape [H, W, C] + :param bboxes: bounding boxes + :param categories: category id + :param instance_semantics: category id corresponding semantic + :param xmlpath: xml file save path + :return: None + """ + doc = Document() + annotation = doc.createElement('annotation') + doc.appendChild(annotation) + folder = doc.createElement('folder') + folder_name = doc.createTextNode('Cityscapes') + folder.appendChild(folder_name) + annotation.appendChild(folder) + + filenamenode = doc.createElement('filename') + filename_name = doc.createTextNode(filename) + filenamenode.appendChild(filename_name) + annotation.appendChild(filenamenode) + + source = doc.createElement('source') + annotation.appendChild(source) + database = doc.createElement('database') + database.appendChild(doc.createTextNode('Cityscapes')) + source.appendChild(database) + + annotation_s = doc.createElement('annotation') + annotation_s.appendChild(doc.createTextNode('PASCAL VOC2007')) + source.appendChild(annotation_s) + + owner = doc.createElement('owner') + name_o = doc.createElement('name') + name_o.appendChild(doc.createTextNode('kinredon')) + owner.appendChild(name_o) + + size = doc.createElement('size') + annotation.appendChild(size) + width = doc.createElement('width') + width.appendChild(doc.createTextNode(str(img_shape[1]))) + height = doc.createElement('height') + height.appendChild(doc.createTextNode(str(img_shape[0]))) + depth = doc.createElement('depth') + depth.appendChild(doc.createTextNode(str(img_shape[2]))) + size.appendChild(width) + size.appendChild(height) + size.appendChild(depth) + + segmented = doc.createElement('segmented') + segmented.appendChild(doc.createTextNode('0')) + annotation.appendChild(segmented) + + for i in range(len(bboxes)): + bbox = bboxes[i] + objects = doc.createElement('object') + annotation.appendChild(objects) + object_name = doc.createElement('name') + object_name.appendChild(doc.createTextNode(instance_semantics[categories[i] - 1])) + objects.appendChild(object_name) + + truncated = doc.createElement('truncated') + truncated.appendChild(doc.createTextNode('0')) + objects.appendChild(truncated) + + difficult = doc.createElement('difficult') + difficult.appendChild(doc.createTextNode('0')) + objects.appendChild(difficult) + bndbox = doc.createElement('bndbox') + objects.appendChild(bndbox) + + xmin = doc.createElement('xmin') + xmin.appendChild(doc.createTextNode(str(bbox[0]))) + bndbox.appendChild(xmin) + ymin = doc.createElement('ymin') + ymin.appendChild(doc.createTextNode(str(bbox[1]))) + bndbox.appendChild(ymin) + xmax = doc.createElement('xmax') + xmax.appendChild(doc.createTextNode(str(bbox[2]))) + bndbox.appendChild(xmax) + ymax = doc.createElement('ymax') + ymax.appendChild(doc.createTextNode(str(bbox[3]))) + bndbox.appendChild(ymax) + + with open(xmlpath, 'w') as f: + f.write(doc.toprettyxml(indent=' ')) \ No newline at end of file