def batch_resize(in_dir, out_dir, ratio=.5):
ensure_dir(out_dir)
for in_f, out_f in files(in_dir=in_dir,
out_dir=out_dir,
in_ext=".jpg",
out_ext=".jpg"):
resize(in_f, out_f, ratio)
for in_f, out_f in files(in_dir=in_dir,
out_dir=out_dir,
in_ext=".png",
out_ext=".png"):
resize(in_f, out_f, ratio)
def batch_undistortion(cal_dir=CAL_DIR,
in_dir=IMG_DIR,
out_dir=UND_DIR,
crop=True):
ensure_dir(out_dir)
cal_f = files(dir=cal_dir, ext=".jpg")[0]
mapx, mapy, roi = undistortion_parameters(cal_f)
for in_f, out_f in files(in_dir=in_dir,
out_dir=out_dir,
in_ext=".jpg",
out_ext=".jpg"):
undistortion(in_f, out_f, mapx, mapy, roi, crop)
def batch_raw2jpg(in_dir=RAW_DIR, out_dir=IMG_DIR):
ensure_dir(out_dir)
for in_f, out_f in files(in_dir=in_dir,
out_dir=out_dir,
in_ext=".nef",
out_ext=".jpg"):
raw2jpg(in_f, out_f)
def prepare_annotator(img_dir=IMG_DIR,
ann_dir=ANN_DIR,
app_dir=APP_DIR,
classes=("background", "plant", "panicle")):
img_out_dir = app_dir + "/data/images"
for f in files(img_out_dir, ".jpg"):
os.remove(f)
for f in files(img_dir, ".jpg"):
shutil.copy(f, img_out_dir)
file_message(img_out_dir + "/" + os.path.basename(f))
ann_out_dir = app_dir + "/data/annotations"
for f in files(ann_out_dir, ".png"):
os.remove(f)
for f in files(ann_dir, ".png"):
shutil.copy(f, ann_out_dir)
file_message(ann_out_dir + "/" + os.path.basename(f))
jsonfile = files(app_dir + "/data", ".json")[0]
with open(jsonfile, "r") as read_file:
data = json.load(read_file)
data["labels"] = classes
data["imageURLs"] = [
"data/images/" + os.path.basename(f)
for f in files(img_out_dir, ".jpg")
]
data["annotationURLs"] = [
"data/annotations/" + os.path.basename(f)
for f in files(ann_out_dir, ".png")
]
with open(jsonfile, "w") as write_file:
json.dump(data, write_file)
file_message(jsonfile)
def batch_threshold_segmentation(limits=([0, 255], [54, 255], [0, 255]),
colour_space="HSV",
in_dir=IMG_DIR,
out_dir=ANN_DIR):
ensure_dir(out_dir)
for in_f, out_f in files(in_dir=in_dir,
out_dir=out_dir,
in_ext=".jpg",
out_ext=".png"):
threshold_segmentation(in_f, out_f, limits, colour_space)
def conda_install_requirements(venv):
"""
Run "pip install -r" on the requirements file.
:Example:
fab --config=config.conf python.conda_install_requirements:venv=bugtrax
"""
# Upload the requirements file.
put(utils.files('requirements', 'base.txt'), utils.home('base.txt'))
put(utils.files('requirements', 'prod.txt'), utils.home('prod.txt'))
# Activate the virtual environment.
activate = '{0}/bin/activate'.format(utils.home('apps', 'miniconda'))
with prefix('source {activate} {venv}'.format(venv=venv, activate=activate)):
run('pip install -r {0}'.format(utils.home('prod.txt')))
# Remove the uploaded files.
with cd(utils.home()):
run('rm {0}'.format(utils.home('base.txt')))
run('rm {0}'.format(utils.home('prod.txt')))
def batch_detect_markers(in_dir=UNDCR_DIR, out_dir=ARUCO_DIR, show=True):
ensure_dir(out_dir)
aruco_dict = cv2.aruco.Dictionary_get(cv2.aruco.DICT_7X7_1000)
aruco_params = cv2.aruco.DetectorParameters_create()
for in_f, out_f in files(in_dir=in_dir,
out_dir=out_dir,
in_ext=".jpg",
out_ext=".pkl"):
detect_markers(in_f,
out_f,
aruco_dict=aruco_dict,
aruco_params=aruco_params,
show=show)
def load_data():
# rumours = {k: {} for k in folders('.')}
data = {}
for rumour, r_location in folders(DATA_LOCATION):
data[rumour] = {}
for thread, t_location in folders(r_location):
try:
replies = files(path.join(t_location, 'replies'))
except StopIteration:
replies = []
# print(rumour, thread, path.join(t_location, 'replies'))
data[rumour][thread] = {
"structure": read(path.join(t_location, 'structure.json')),
"source": tweet(path.join(t_location, 'source-tweet', thread + '.json')),
"replies": {id[:-5]: tweet(f) for id, f in replies}
}
write('data/data.json', data)
return data
import os
import pandas
from lxml import etree
from decimal import Decimal
from utils import files, int_file, pairs
_keys = map(int_file, files('ufrawkey'))
_frames = dict(map(int_file, files('cr2')))
exposure_keys = list()
basemanualcurve_keys = list()
manualcurve_keys = list()
def curve_anchors(tree, curve_name):
anchors = list()
for anchorxy in tree.find(curve_name).iterchildren():
X, Y = anchorxy.text.split()
anchors.extend([Decimal(X), Decimal(Y)])
return anchors
for int_, filename in sorted(_keys):
with open(filename) as keyframe_file:
contents = keyframe_file.read()
parser = etree.XMLParser()
parser.feed(contents)
tree = parser.close()
exposure_keys.append(
(int(int_), Decimal(tree.find('Exposure').text)))
def main():
args = get_parser()
# source faces
srcFaces = tqdm(files(args.srcFacePath, ['.jpg']))
# real faces database
#ds = image2pilBatch(files(args.faceDatabase, ['.jpg']))
# face detector
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor(args.shapePredictor)
for i, srcFace in enumerate(srcFaces):
# load bgr
try:
srcFaceBgr = cv2.imread(srcFace)
except:
tqdm.write(f'Fail loading: {srcFace}')
continue
# detect landmarks
srcLms = get_landmarks(detector, predictor,
cv2.cvtColor(srcFaceBgr, cv2.COLOR_BGR2RGB))
if srcLms is None:
tqdm.write(f'No face: {srcFace}')
continue
# find first face whose landmarks are close enough in real face database
targetRgb = find_one_neighbor(detector, predictor, srcFace, srcLms,
files(args.faceDatabase, ['.jpg']),
args.threshold)
if targetRgb is None: # if not found
tqdm.write(f'No Match: {srcFace}')
continue
# if found
targetBgr = cv2.cvtColor(targetRgb, cv2.COLOR_RGB2BGR)
hullMask = convex_hull(
srcFaceBgr.shape,
srcLms) # size (h, w, c) mask of face convex hull
# generate random deform
anchors, deformedAnchors = random_deform(hullMask.shape[:2], 4, 4)
# piecewise affine transform and blur
warped = piecewise_affine_transform(
hullMask, anchors, deformedAnchors) # size (h, w) warped mask
blured = cv2.GaussianBlur(warped, (5, 5), 3)
# swap
left, up, right, bot = min(srcLms[:, 0]), min(srcLms[:, 1]), max(
srcLms[:, 0]), max(srcLms[:, 1])
targetBgrT = color_transfer(srcFaceBgr[up:bot, left:right, :],
targetBgr)
resultantFace = forge(srcFaceBgr, targetBgrT, blured) # forged face
# save face images
cv2.imwrite(f'./dump/mask_{i}.jpg', hullMask)
cv2.imwrite(f'./dump/deformed_{i}.jpg', warped * 255)
cv2.imwrite(f'./dump/blured_{i}.jpg', blured * 255)
cv2.imwrite(f'./dump/src_{i}.jpg', srcFaceBgr)
cv2.imwrite(f'./dump/target_{i}.jpg', targetBgr)
cv2.imwrite(f'./dump/target_T_{i}.jpg', targetBgrT)
cv2.imwrite(f'./dump/forge_{i}.jpg', resultantFace)
def _needs_remove(self, directory):
return any((os.path.splitext(item)[1] in self.supported_suffixes
and self._has_tags(item)) for item in files(directory))
def _needs_add(self, directory, force):
return any((os.path.splitext(item)[1] in self.supported_suffixes and (
not self._has_tags(item) or force)) for item in files(directory))
def main():
args = get_parser()
import dlib
# source faces
srcFaces = tqdm(files(args.srcFacePath, ['.jpg']))
# real faces database
# ds = image2pilBatch(files(args.faceDatabase, ['.jpg']))
# face detector
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor(args.shapePredictor)
for i, srcFace in enumerate(srcFaces):
# load bgr
try:
srcFaceBgr = cv2.imread(srcFace)
except:
tqdm.write(f'Fail loading: {srcFace}')
continue
# detect landmarks
srcLms = get_landmarks(detector, predictor,
cv2.cvtColor(srcFaceBgr, cv2.COLOR_BGR2RGB))
if srcLms is None:
tqdm.write(f'No face: {srcFace}')
continue
# find first face whose landmarks are close enough in real face database
targetRgb = find_one_neighbor(detector, predictor, srcFace, srcLms,
files(args.faceDatabase, ['.jpg']),
args.threshold)
if targetRgb is None: # if not found
tqdm.write(f'No Match: {srcFace}')
continue
# if found
# 产生凸包
targetBgr = cv2.cvtColor(targetRgb, cv2.COLOR_RGB2BGR)
hullMask = convex_hull(
srcFaceBgr.shape, srcLms
) # size (h, w, c) mask of face convex hull, uint8 [0, 255] # TODO: 去掉channel减少参数
# 产生随机变形
anchors, deformedAnchors = random_deform(hullMask.shape[:2], 4,
4) # 随机变形存在问题：需要取变形结果与原图的交集。
# 分段仿射变换
warped = piecewise_affine_transform(
hullMask, anchors,
deformedAnchors) # size (h, w, c) warped mask float64 [0, 1.0]
# 将 warped 区域限制在人脸范围内，避免背景的影响
warped *= (hullMask / hullMask.max())
# 高斯模糊
blured = cv2.GaussianBlur(warped, (5, 5), 3)
# 颜色矫正，迁移高斯模糊后blured mask区域的颜色，并对颜色纠正的blured区域人脸+原背景作为融合图片
left, up, right, bot = get_roi(warped) # 获取 warped 区域
src = (srcFaceBgr[up:bot, left:right, :]).astype(np.uint8)
tgt = (targetBgr[up:bot, left:right, :]).astype(np.uint8)
# pdb.set_trace()
targetBgrT = color_transfer(src, tgt)
cv2.imwrite(f'results/transfer/src.jpg', src)
cv2.imwrite(f'results/transfer/tgt.jpg', tgt)
cv2.imwrite(f'results/transfer/tgtrans.jpg', targetBgrT)
targetBgr_T = targetBgr * 1 # 开辟新内存空间
targetBgr_T[up:bot, left:right, :] = targetBgrT # 将色彩迁移的部分转移到原图片
# 融合
resultantFace = forge(srcFaceBgr, targetBgr_T, blured) # forged face
# 混合边界
resultantBounding = get_bounding(blured)
# save face images
cv2.imwrite(f'./results/mask_{i}.jpg', hullMask)
cv2.imwrite(f'./results/deformed_{i}.jpg', warped * 255)
cv2.imwrite(f'./results/blured_{i}.jpg', blured * 255)
cv2.imwrite(f'./results/src_{i}.jpg', srcFaceBgr)
cv2.imwrite(f'./results/target_{i}.jpg', targetBgr)
cv2.imwrite(f'./results/target_T_{i}.jpg', targetBgrT)
cv2.imwrite(f'./results/forge_{i}.jpg', resultantFace)
cv2.imwrite(f'./results/bounding_{i}.jpg', resultantBounding * 255)
import os
from multiprocessing import Pool
from subprocess import check_call
from utils import files
def render_jpg(R):
return R
if __name__ == '__main__':
pool = Pool(8)
raw_files = files('cr2')
num = len(raw_files)
for pos, ret in enumerate(pool.imap_unordered(render_jpg, raw_files)):
pos += 1
fmt_string = "{0}/{1} ({2:.3f}%) done"
print(fmt_string.format(pos + 1, num, ((pos + 1) / float(num)) * 100))
pool.close()
pool.join()
