def draw_bbox_on_tiff(chip_path, coords, chips, classes, save_path):
#Load an image
#path = '/home/ubuntu/anyan/harvey_data/converted_sample_tiff/'
# big tiff name: chip name
# init to {big_tiff_name : []}
#big_tiff_dict = dict((k, []) for k in big_tiff_set)
big_tiff_dict = dict()
fnames = glob.glob(chip_path + "*.tif")
# i = 0
for f in fnames:
chip_name = f.split('/')[-1]
chip_big_tiff_id_list = chip_name.split('_')[1:3]
chip_big_tiff_id = '_'.join(chip_big_tiff_id_list)
print(chip_big_tiff_id)
if chip_big_tiff_id not in set(big_tiff_dict.keys()):
big_tiff_dict[chip_big_tiff_id] = list()
else:
if len(big_tiff_dict[chip_big_tiff_id]) < 5:
big_tiff_dict[chip_big_tiff_id].append(chip_name)
arr = wv.get_image(f)
# print(arr.shape)
# plt.figure(figsize=(10,10))
# plt.axis('off')
# plt.imshow(arr)
coords_chip = coords[chips==chip_name]
#print(coords_chip.shape)
if coords_chip.shape[0] == 0:
continue
classes_chip = classes[chips==chip_name].astype(np.int64)
# #We can chip the image into 500x500 chips
# c_img, c_box, c_cls = wv.chip_image(img = arr, coords= coords, classes=classes, shape=(500,500))
# print("Num Chips: %d" % c_img.shape[0])
labelled = aug.draw_bboxes(arr,coords_chip[classes_chip ==1])
print(chip_name)
# plt.figure(figsize=(15,15))
# plt.axis('off')
# plt.imshow(labelled)
subdir_name = save_path + chip_big_tiff_id
if os.path.isdir(subdir_name):
save_name = subdir_name +'/' + chip_name + '.png'
print(save_name)
labelled.save(save_name)
else:
os.mkdir(subdir_name)
save_name = subdir_name +'/' + chip_name + '.png'
print(save_name)
labelled.save(save_name)
else:
continue
def parseChip(self,
c_img,
c_box,
c_cls,
c_ids,
img_num,
c_dir,
res_out=30,
showImg=False):
# Parses chips, saves chip image, and also saves corresponding labels
fnames = []
outputImgDir = "{}labels/".format(c_dir)
outputLabelDir = "{}images/".format(c_dir)
self.checkDir(outputImgDir)
self.checkDir(outputLabelDir)
images = []
sboxes, sclasses, simgs = [], [], []
for c_idx in range(c_img.shape[0]):
c_name = "{:06}_{:02}".format(int(img_num), c_idx)
sbox,scls,simg,sid = \
c_box[c_idx],c_cls[c_idx],c_img[c_idx], c_ids[c_idx]
#print(sid)
if showImg:
labelled = aug.draw_bboxes(simg, sbox)
plt.figure(figsize=(10, 10))
plt.imshow(labelled)
#print(c_box[:,4])
plt.title(str(c_name) + " " + str(simg.shape))
plt.axis('off')
break
# Change chip into darknet format, and save
result = self.toDarknetFmt(sbox, scls, simg, sid)
ff_l = "{}labels/{}.txt".format(c_dir, c_name)
np.savetxt(ff_l, result, fmt='%i %1.6f %1.6f %1.6f %1.6f %i')
# Save image to specified dir
ff_i = "{}images/{}.jpg".format(c_dir, c_name)
Image.fromarray(simg).save(ff_i)
# Append file name to list
fnames.append("{}images/{}.jpg".format(c_dir, c_name))
pass
return fnames
continue
'''
if local_class1 > 0:
num_class1_chip += 1
if local_class2 > 0:
num_class2_chip += 1
#else:
train_writer.write(tf_example.SerializeToString())
train_chips += 1
if SAVE_IMAGES and idx % 5 == 0:
# debug: changed save dir
#aug.draw_bboxes(image, new_coords[new_classes ==1]).save('./harvey_ms_img_inspect_train_2class_noclean/img_%s_%s.png'%(name,str(idx)))
aug.draw_bboxes(image, new_coords).save(
'./tomnod_2class_train_inspect/img_%s_%s.png' %
(name, str(idx)))
ind_chips += 1
# debug
# store the training and validation images with bboxes for inspection
'''
if SAVE_IMAGES:
# debug: changed save dir
aug.draw_bboxes(image, new_coords).save('./harvey_img_inspect/img_%s_%s.png'%(name,str(idx)))
'''
#Make augmentation probability proportional to chip size. Lower chip size = less chance.
#This makes the chip-size imbalance less severe.
#prob = np.random.randint(0,np.max(res))
(name, extra, it[0]))
if len(nb) > 0:
tf_example = tfr.to_tf_example(
newimg, nb, classes_final[idx])
#Don't count augmented chips for chip indices
if idx < split_ind:
test_writer.write(
tf_example.SerializeToString())
test_chips += 1
else:
train_writer.write(
tf_example.SerializeToString())
train_chips += 1
else:
if SAVE_IMAGES:
aug.draw_bboxes(newimg, nb).save(
'process/img_nobox_%s_%s_%s.png' %
(name, extra, it[0]))
if res_ind == 0:
max_chips_per_res = int(ind_chips * 1.5)
logging.info("Max chips per resolution: %s " % max_chips_per_res)
logging.info("Tot Box: %d" % tot_box)
logging.info("Chips: %d" % ind_chips)
logging.info("saved: %d train chips" % train_chips)
logging.info("saved: %d test chips" % test_chips)
train_writer.close()
test_writer.close()
if local_class1 > 0:
num_class1_chip += 1
if local_class2 > 0:
num_class2_chip += 1
#if idx < split_ind:
test_writer.write(tf_example.SerializeToString())
test_chips += 1
if SAVE_IMAGES and idx % 5 == 0:
# debug: changed save dir
#debug
# draw only DAMAGED buildings
#aug.draw_bboxes(image, new_coords[new_classes ==1]).save('./harvey_ms_img_inspect_val_2class_noclean/img_%s_%s.png'%(name,str(idx)))
aug.draw_bboxes(image, new_coords).save(
'./tomnod_valtest_2class_inspect/img_%s_%s.png' %
(name, str(idx)))
#else:
# train_writer.write(tf_example.SerializeToString())
# train_chips += 1
# if SAVE_IMAGES:
# debug: changed save dir
#aug.draw_bboxes(image, new_coords[new_classes ==1]).save('./harvey_ms_img_inspect_train_2class_noclean/img_%s_%s.png'%(name,str(idx)))
# aug.draw_bboxes(image, new_coords).save('./harvey_ms_img_inspect_train_2class_noclean/img_%s_%s.png'%(name,str(idx)))
ind_chips += 1
# debug
# store the training and validation images with bboxes for inspection
'''
train_writer.write(
tf_example_aug.SerializeToString())
num_aug = num_aug + 1
train_chips += 1
num_aug_per_class[
class_id] = num_aug_per_class[
class_id] + 1
# num_aug_this_class=num_aug_this_class + 1
# debug
if aug_idx % 10 == 0 and SAVE_IMAGES:
# debug: changed save dir
aug_image = (aug_image).astype(
np.uint8)
aug.draw_bboxes(
aug_image, boxes_aug[aug_idx]
).save(
'./expand_aug_random_256/img_aug_%s_%s_%s_%s.png'
% (name, str(idx), str(aug_idx),
str(class_id)))
# debug
print('augmenting class: ', class_id)
print('number of augmentation: ', num_aug)
#num_aug_per_class[class_id] = num_aug_this_class
# it: iterator for different resolutions
# start to augment the rest
if AUGMENT and prob < it[0] and MINOR_CLASS_FLAG == False:
for extra in range(3):
center = np.array([
int(image.shape[0] / 2),
int(image.shape[1] / 2)
if local_class1 > 0:
num_class1_chip += 1
if local_class2 > 0:
num_class2_chip += 1
#if idx < split_ind:
test_writer.write(tf_example.SerializeToString())
test_chips += 1
if SAVE_IMAGES and idx % 5 == 0:
# debug: changed save dir
#debug
# draw only DAMAGED buildings
#aug.draw_bboxes(image, new_coords[new_classes ==1]).save('./harvey_ms_img_inspect_val_2class_noclean/img_%s_%s.png'%(name,str(idx)))
aug.draw_bboxes(image, new_coords[new_classes == 1]).save(
'./tomnod_valtest_1class_inspect/img_%s_%s.png' %
(name, str(idx)))
#else:
# train_writer.write(tf_example.SerializeToString())
# train_chips += 1
# if SAVE_IMAGES:
# debug: changed save dir
#aug.draw_bboxes(image, new_coords[new_classes ==1]).save('./harvey_ms_img_inspect_train_2class_noclean/img_%s_%s.png'%(name,str(idx)))
# aug.draw_bboxes(image, new_coords).save('./harvey_ms_img_inspect_train_2class_noclean/img_%s_%s.png'%(name,str(idx)))
ind_chips += 1
# debug
# store the training and validation images with bboxes for inspection
'''
continue
'''
if local_class1 > 0:
num_class1_chip += 1
if local_class2 > 0:
num_class2_chip += 1
#else:
train_writer.write(tf_example.SerializeToString())
train_chips += 1
if SAVE_IMAGES and idx % 5 == 0:
# debug: changed save dir
#aug.draw_bboxes(image, new_coords[new_classes ==1]).save('./harvey_ms_img_inspect_train_2class_noclean/img_%s_%s.png'%(name,str(idx)))
aug.draw_bboxes(image, new_coords).save(
'./noaa_train_2class_inspect/img_%s_%s.png' %
(name, str(idx)))
ind_chips += 1
# debug
# store the training and validation images with bboxes for inspection
'''
if SAVE_IMAGES:
# debug: changed save dir
aug.draw_bboxes(image, new_coords).save('./harvey_img_inspect/img_%s_%s.png'%(name,str(idx)))
'''
#Make augmentation probability proportional to chip size. Lower chip size = less chance.
#This makes the chip-size imbalance less severe.
#prob = np.random.randint(0,np.max(res))
# ...
# ])
# 1: array([ [...], [...], ... ])
# ...
# }
#
# c_cls = dictionary keyed by integer and values are arrays of ints which are class labels.
# i.e. c_cls = {
# 0: [10, 15, 21, 29, 71, 20]
# 1: [30, 55, 22, 19, 61, 30]
# ...
# }
# The keys of the dictionaries match up with len(c_img). The details for chip
# c_img[0] are stored in c_box[0] and c_cls[0]. So, len(c_img) == len(c_box) == len(c_cls)
c_img, c_box, c_cls = wv.chip_image(img=arr,
coords=coords,
classes=classes,
shape=(700, 700))
# Augment the data
ind = np.random.choice(range(c_img.shape[0]))
center = (int(c_img[ind].shape[0] / 2), int(c_img[ind].shape[1] / 2))
for deg in range(0, 360, 45):
rot_im, rot_boxes = aug.rotate_image_and_boxes(c_img[ind], deg, center,
c_box[ind])
pim = aug.draw_bboxes(rot_im, rot_boxes)
plt.imshow(pim)
plt.title("deg: {}".format(deg))
plt.show()
def exportChipImages(self,
image_paths,
c_dir,
set_str,
res_out=30,
showImg=False,
shape=(600, 600),
chipImage=False):
fnames = []
#image_paths = sorted(image_paths)
for img_pth in image_paths:
try:
img_pth = self.input_dir + 'train_images/' + img_pth
img_name = img_pth.split("/")[-1]
img_num = img_name.split(".")[0]
arr = wv.get_image(img_pth)
chip_coords = self.coords[self.chips == img_name]
chip_classes = self.classes[self.chips == img_name].astype(
np.int64)
chip_coords,chip_classes = \
self.filterClasses(chip_coords,chip_classes,self.grouped_classes)
# Code for downsampling the image
scale = self.res_native / res_out
if res_out != 30:
arr = self.downsampleImage(arr, res_out=res_out)
chip_coords[:, :4] = chip_coords[:, :4] * scale
# Chip the tif image into tiles
c_img, c_box, c_cls, c_ids = chip_image(img=arr,
coords=chip_coords,
classes=chip_classes,
shape=shape,
chipImage=chipImage)
if showImg:
result = []
for key in c_cls.keys():
result.extend(c_cls[key])
print("number of classes: {}".format(len(result)))
for i, img in enumerate(c_img[:5]):
labelled = aug.draw_bboxes(c_img[i], c_box[i])
plt.imshow(labelled)
plt.axis('off')
plt.show()
pass
break
pass
c_fnames = self.parseChip(c_img, c_box, c_cls, c_ids, img_num,
c_dir, res_out)
fnames.extend(c_fnames)
except FileNotFoundError as e:
print(e)
pass
pass
lines = sorted(fnames)
if not os.path.exists(self.output_dir):
os.makedirs(self.output_dir)
pass
outputTxtPath = self.output_dir + "xview_img_{}.txt".format(set_str)
print(outputTxtPath)
if os.path.exists(outputTxtPath):
os.remove(outputTxtPath)
pass
with open(outputTxtPath, mode='w', encoding='utf-8') as myfile:
myfile.write('\n'.join(lines))
pass
