def __getitem__(self, item):
image, landmarks = super(BBoxDataset, self).__getitem__(item)
bbox_path = self.bboxes[item]
try:
bbox = utils.read_bbox(bbox_path)
except:
bbox = [0, 0, image.shape[1] - 1, image.shape[0] - 1]
minx, miny, maxx, maxy = bbox
if self.phase == 'train':
left = min(minx, self.max_jitter)
right = min(image.shape[1] - maxx - 1, self.max_jitter)
up = min(miny, self.max_jitter)
down = min(image.shape[0] - maxy - 1, self.max_jitter)
dh = np.random.randint(-up, down + 1, 1)
dv = np.random.randint(-left, right + 1, 1)
bbox[0::2] += dv
bbox[1::2] += dh
minx, miny, maxx, maxy = bbox
landmarks = utils.norm_landmarks(landmarks, bbox)
image = image[miny:maxy + 1, minx:maxx + 1, :]
image = cv2.resize(image, self.shape)
if self.phase == 'train':
image, landmarks = utils.random_flip(image, landmarks, 0.5)
image = utils.random_gamma_trans(image,
np.random.uniform(0.8, 1.2, 1))
image = utils.random_color(image)
image = np.transpose(image, (2, 0, 1)).astype(np.float32)
return image, np.reshape(landmarks, (-1))
def save_pic_and_lmk(kk):
shuru = kk
if shuru == 0:
e = 1 #return 0 输入为0,说明只有一张图
else:
lmk_dir = '/home/zhang/correctdata/data/landmark/' #注意lmk文件的后缀为.txt
pic_dir = '/home/zhang/correctdata/data/picture/'
bbox_dir = '/home/zhang/correctdata/bbox/'
length = len(shuru)
for i in range(length):
if i % 2 == 0:
filename1 = pic_dir + shuru[i]
'''filename2 = pic_dir+shuru[i+1]'''
lmk1 = utils.read_mat(lmk_dir + shuru[i + 1] + '.txt')
lmk2 = utils.read_mat(lmk_dir + shuru[i] + '.txt')
lmk = 0.5 * lmk1 + 0.5 * lmk2
image = plt.imread(filename1)
#图一的landmark归一化
bbox1 = utils.read_bbox(bbox_dir + shuru[i + 1] + '.rect')
lmk1 = utils.norm_landmarks(lmk1, bbox1)
image = image[bbox1[1]:bbox1[3], bbox1[0]:bbox1[2]]
#图一图二的插值图的landmark归一化
bbox2 = utils.read_bbox(bbox_dir + shuru[i] + '.rect')
lmk = utils.norm_landmarks(lmk, bbox2)
out = pwa.pwa(image.copy(), lmk1.copy(), lmk.copy(),
(256, 256))
# 图片与lmk保存
utils.save_landmarks(
lmk * 255, '/home/zhang/ming/save_for_lmk2/' +
shuru[i][:-5] + "%d" % (200 + i) + '.jpg' + '.txt')
io.imsave(
'/home/zhang/ming/save_for_pic2/' + shuru[i][:-5] + "%d" %
(200 + i) + '.jpg', out)
def __getitem__(self, i):
img_path = self.images[i]
bbox_path = self.bboxes[i]
landmark_path = self.landmarks[i]
bbox = utils.read_bbox(bbox_path)
landmarks = utils.read_mat(landmark_path)
landmarks = utils.norm_landmarks(landmarks, bbox)
image = cv2.imread(img_path)
# minx, miny, maxx, maxy = bbox
# image = image[miny:maxy+1, minx:maxx+1, :]
# image = cv2.resize(image, self.shape)
# cv2.imshow("t", image)
# cv2.waitKey(0)
image = np.transpose(image, (2, 0, 1)).astype(np.float32)
#print('origin', origin_landmarks)
return image, np.reshape(landmarks,
(-1)), np.reshape(np.array(bbox), (2, 2))
def main():
if not os.path.exists('cache'):
os.system('mkdir cache')
# 我偷懒了，所以最好不要写成'/home/yqi/data/icme/'
root_dir = '/data/icme'
lamdmark_dir = os.path.join(root_dir, 'data/landmark')
image_dir = os.path.join(root_dir, 'data/picture')
bbox_dir = os.path.join(root_dir, 'bbox')
try:
filenames = joblib.load('cache/filenames.pkl')
norm_landmarks = joblib.load('cache/norm_landmarks.pkl')
mean_landmarks = joblib.load('cache/mean_landmarks.pkl')
bboxes = joblib.load('cache/bboxes.pkl')
split = joblib.load('cache/split.pkl')
except:
filenames = os.listdir(image_dir)
norm_landmarks = []
bboxes = []
split = {}
for filename in filenames:
id = get_id(filename)
if np.random.uniform(0, 1) < 0.8:
split[id] = 'train'
landmark_path = os.path.join(lamdmark_dir, filename + '.txt')
bbox_path = os.path.join(bbox_dir, filename + '.rect')
bbox = utils.read_bbox(bbox_path)
landmarks = utils.read_mat(landmark_path)
landmarks = utils.norm_landmarks(landmarks, bbox)
norm_landmarks.append(landmarks)
else:
split[id] = 'valid'
bbox_path = os.path.join(bbox_dir, filename + '.rect')
bbox = utils.read_bbox(bbox_path)
bboxes.append(bbox)
norm_landmarks = np.stack(norm_landmarks, axis=0)
mean_landmarks = np.mean(norm_landmarks, axis=0)
joblib.dump(norm_landmarks, 'cache/norm_landmarks.pkl', compress=3)
joblib.dump(mean_landmarks, 'cache/mean_landmarks.pkl', compress=3)
joblib.dump(filenames, 'cache/filenames.pkl', compress=3)
joblib.dump(bboxes, 'cache/bboxes.pkl', compress=3)
joblib.dump(split, 'cache/split.pkl', compress=3)
# for i in range(106):
# plt.scatter(mean_landmarks[i, 0], mean_landmarks[i, 1])
# plt.show()
try:
transform_matrix = joblib.load('cache/transform_matrix.pkl')
aligned = joblib.load('cache/aligned.pkl')
except:
transform_matrix = []
aligned = []
i = -1
for filename in filenames:
if split[get_id(filename)] == 'valid':
continue
i += 1
curr = norm_landmarks[i, :]
one = np.ones(shape=(106, 1))
curr = np.concatenate((curr, one), axis=1)
t = procrustes(curr, mean_landmarks)
transform_matrix.append(t)
aligned.append(np.reshape(curr@t, (-1)))
joblib.dump(transform_matrix, 'cache/transform_matrix.pkl', compress=3)
joblib.dump(aligned, 'cache/aligned.pkl', compress=3)
temp = (aligned - np.mean(aligned, axis=0))
covariance = 1.0 / len(aligned) * temp.T.dot(temp)
U, S, V = np.linalg.svd(covariance)
joblib.dump(U, 'cache/u.pkl', compress=3)
pc = temp.dot(U[:, 0])
plt.hist(pc,bins=11)
plt.show()
for i, filename in enumerate(filenames):
img_path = os.path.join(image_dir, filename)
if pc[i] > 0.793:
n = '1'
elif pc[i] > 0.615:
n = '2'
elif pc[i] > 0.44:
n = '3'
elif pc[i] > 0.26:
n = '4'
elif pc[i] > 0.087:
n = '5'
elif pc[i] > -0.0913:
n = '6'
elif pc[i] > -0.264:
n = '7'
elif pc[i] > -0.448:
n = '8'
elif pc[i] > -0.62:
n = '9'
elif pc[i] > -0.79:
n = '10'
else:
n = '11'
id = get_id(filename)
cmd = 'ln -s %s %s/%s/%s/%s' % (img_path, root_dir, split[id], n, filename)
os.system(cmd)
def main():
def get_id(name):
t = name.split('_')[0:2]
return t[0] + t[1]
root_dir = 'D:\icmedata\correctdata\\'
lamdmark_dir = os.path.join(root_dir, 'data/landmark')
image_dir = os.path.join(root_dir, 'data/picture')
bbox_dir = os.path.join(root_dir, 'bbox')
filenames = os.listdir(image_dir)
norm_landmarks = []
bboxes = []
split = {}
for filename in filenames:
id = get_id(filename)
if np.random.uniform(0, 1) < 0.8:
split[id] = 'train'
else:
split[id] = 'valid'
landmark_path = os.path.join(lamdmark_dir, filename + '.txt')
bbox_path = os.path.join(bbox_dir, filename + '.rect')
bbox = utils.read_bbox(bbox_path)
landmarks = utils.read_mat(landmark_path)
landmarks = utils.norm_landmarks(landmarks, bbox)
norm_landmarks.append(landmarks)
bboxes.append(bbox)
norm_landmarks = np.stack(norm_landmarks, axis=0)
mean_landmarks = np.mean(norm_landmarks, axis=0)
# for i in range(106):
# plt.scatter(mean_landmarks[i, 0], mean_landmarks[i, 1])
# plt.show()
target=[]
for i, filename in enumerate(filenames):
curr = norm_landmarks[i, :]
y = curr[1::2]
y_max = np.max(y)
y_min = np.min(y)
x = curr[::2]
x_max = np.max(x)
x_min = np.min(x)
chang = x_max - x_min
kuan = y_max - y_min
Slandmark = chang * kuan
#print((Slandmark))
#print(bboxes[i])
# bbox_tempt = np.array(bboxes)
# Sbbox = (bbox_tempt[:, 2] - bbox_tempt[:, 0]) * (bbox_tempt[:, 3] - bbox_tempt[:, 1])
# print(Sbbox[i])
#landmark就是基于bbox做的归一化在untils。norm——landmark所以就不用求Sbbox
target.append(Slandmark)
draw_hist(target, 'acreage title', 'SL/SB', 'amount', 0, 1, 0, 3000)
for i, filename in enumerate(filenames):
img_path = os.path.join(image_dir, filename)
if target[i] > 0.8:
n = 's1'
elif target[i] > 0.75:
n = 's2'
elif target[i] > 0.7:
n = 's3'
elif target[i] > 0.64:
n = 's4'
elif target[i] > 0.6:
n = 's5'
elif target[i] > 0.54:
n = 's6'
elif target[i] > 0.5:
n = 's7'
else:
n = 's8'
id = get_id(filename)
cmd = 'ln -s %s %s/%s/%s/%s' % (img_path, root_dir, split[id], n, filename)
os.system(cmd)
tform = PiecewiseAffineTransform()
tform.estimate(dst, src)
# out_rows ,out_cols = shape
out_rows = image.shape[0]
out_cols = image.shape[1]
out = warp(image, tform, output_shape=(out_rows, out_cols))
return out
if __name__ == '__main__':
import matplotlib.pyplot as plt
image = plt.imread('/data/icme/data/picture/AFW_134212_1_0.jpg')
bbox = utils.read_bbox('/data/icme/bbox/AFW_134212_1_0.jpg.rect')
src = utils.read_mat('/data/icme/data/landmark/AFW_134212_1_0.jpg.txt')
src = utils.norm_landmarks(src, bbox)
image = image[bbox[1]:bbox[3], bbox[0]:bbox[2]]
bbox = utils.read_bbox('/data/icme/bbox/AFW_134212_1_3.jpg.rect')
dst = utils.read_mat('/data/icme/data/landmark/AFW_134212_1_3.jpg.txt')
dst = utils.norm_landmarks(dst, bbox)
out = pwa(image, src, dst, (128, 128))
plt.subplot(1, 2, 1)
plt.imshow(image)
plt.subplot(1, 2, 2)
plt.imshow(out)
# plt.plot(tform.inverse(src)[:, 0], tform.inverse(src)[:, 1], '.b')
plt.axis((0, 128, 128, 0))
plt.show()
import numpy as np
import cv2
from data import utils
imgdir = '/data/icme/data/picture/AFW_5452623_1_5.jpg'
landmarks = utils.read_mat('/data/icme/data/landmark/AFW_5452623_1_5.jpg.txt')
bbox = utils.read_bbox('/data/icme/bbox/AFW_5452623_1_5.jpg.rect')
img = cv2.imread(imgdir)
minx, miny, maxx, maxy = bbox
img = img[miny:maxy+1, minx:maxx+1, :]
landmarks = utils.norm_landmarks(landmarks, bbox)
img, landmarks = utils.random_flip(img, landmarks, 1)
img = np.transpose(img, (2, 0, 1))
img = utils.draw_landmarks(img, landmarks, (255, 255, 255))
img = np.transpose(img, (1, 2, 0))
cv2.imshow('', img)
cv2.waitKey(0)