def build_dataset(path_list, box_path_list, is_train=True):
filenames = []
landmarks = []
faceboxes = []
for i in range(len(path_list)):
filenames_dir = glob.glob(path_list[i] + "*.jpg")
filenames_dir += glob.glob(path_list[i] + "*.png")
box_dict = pickle.load(open(box_path_list[i], 'rb'))
for j in range(len(filenames_dir)):
filenames.append(filenames_dir[j])
pts_filename = filenames_dir[j][:-3] + "pts"
landmarks.append(loadFromPts(pts_filename))
basename = os.path.basename(filenames_dir[j])
faceboxes.append(box_dict[basename])
if is_train:
with open('data/train_filenames.pkl', 'wb') as f:
pickle.dump(filenames, f, pickle.HIGHEST_PROTOCOL)
with open('data/train_faceboxes.pkl', 'wb') as f:
pickle.dump(faceboxes, f, pickle.HIGHEST_PROTOCOL)
with open('data/train_landmarks.pkl', 'wb') as f:
pickle.dump(landmarks, f, pickle.HIGHEST_PROTOCOL)
else:
with open('data/valid_filenames.pkl', 'wb') as f:
pickle.dump(filenames, f, pickle.HIGHEST_PROTOCOL)
with open('data/valid_faceboxes.pkl', 'wb') as f:
pickle.dump(faceboxes, f, pickle.HIGHEST_PROTOCOL)
with open('data/valid_landmarks.pkl', 'wb') as f:
pickle.dump(landmarks, f, pickle.HIGHEST_PROTOCOL)
def PrepareData(self, imageDirs, boundingBoxFiles, meanShape, startIdx,
nImgs, mirrorFlag):
filenames = [] # pic文件名list
landmarks = [] # 坐标点文件名list
boundingBoxes = []
for i in range(len(imageDirs)): # 训练数据来自lfpw，helen，afw这3个数据集，文件夹长度为3。
filenamesInDir = glob.glob(imageDirs[i] + "*.jpg")
filenamesInDir += glob.glob(
imageDirs[i] + "*.png"
) # 打开第一个文件夹的所有jpg和png图片，文件路径 '../data/images/lfpw/trainset\\image_0020.png'
if boundingBoxFiles is not None: #
boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'),
encoding='latin1')
for j in range(len(filenamesInDir)):
filenames.append(
filenamesInDir[j]) # 将每1个pic的路径，加入filenames所维护的list中
ptsFilename = filenamesInDir[j][:-3] + "pts"
landmarks.append(
utils.loadFromPts(ptsFilename)
) # 将每个图片标注信息的pts路径，写入list：'../data/images/lfpw/trainset\\image_0001.pts'
if boundingBoxFiles is not None:
basename = path.basename(filenamesInDir[j])
boundingBoxes.append(boundingBoxDict[basename])
filenames = filenames[startIdx:startIdx + nImgs]
landmarks = landmarks[startIdx:startIdx + nImgs]
boundingBoxes = boundingBoxes[startIdx:startIdx + nImgs]
mirrorList = [False for i in range(nImgs)]
if mirrorFlag:
mirrorList = mirrorList + [True for i in range(nImgs)]
filenames = np.concatenate((filenames, filenames))
landmarks = np.vstack((landmarks, landmarks)) # 6096个pic对应的特征点信息
boundingBoxes = np.vstack((boundingBoxes, boundingBoxes))
self.origLandmarks = landmarks # 将读取到的信息，写到对应的mageServer的属性成员上
self.filenames = filenames
self.mirrors = mirrorList
self.meanShape = meanShape
self.boundingBoxes = boundingBoxes
def PrepareData(self, imageDirs, boundingBoxFiles, meanShape, startIdx,
nImgs, mirrorFlag):
filenames = [
] # 이 목록은 filenamesInDir과 완전히 동일하므로 제거 할 수 있습니다. <<< 此list和filenamesInDir完全一样，可以去除
landmarks = [] # 이 리스트는 특징점 좌표를 저장하는 데 사용됩니다. <<< 此list用于存储特征点坐标
boundingBoxes = [] # 이 목록은 bbx를 저장하는 데 사용됩니다. <<< 此list用于存储bbx
# import pdb; pdb.set_trace()
for i in range(len(imageDirs)):
filenamesInDir = glob.glob(imageDirs[i] + "*.jpg")
filenamesInDir += glob.glob(imageDirs[i] + "*.png")
# import pdb; pdb.set_trace()
if boundingBoxFiles is not None:
boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'))
for j in range(len(filenamesInDir)):
filenames.append(filenamesInDir[j])
ptsFilename = filenamesInDir[j][:-3] + "pts"
landmarks.append(utils.loadFromPts(ptsFilename))
if boundingBoxFiles is not None:
basename = path.basename(filenamesInDir[j])
boundingBoxes.append(boundingBoxDict[basename])
filenames = filenames[startIdx:startIdx + nImgs]
landmarks = landmarks[startIdx:startIdx + nImgs]
boundingBoxes = boundingBoxes[startIdx:startIdx + nImgs]
mirrorList = [False for i in range(nImgs)]
if mirrorFlag:
mirrorList = mirrorList + [True for i in range(nImgs)]
filenames = np.concatenate((filenames, filenames))
landmarks = np.vstack((landmarks, landmarks))
boundingBoxes = np.vstack((boundingBoxes, boundingBoxes))
self.origLandmarks = landmarks
self.filenames = filenames
self.mirrors = mirrorList
self.meanShape = meanShape
self.boundingBoxes = boundingBoxes
def PrepareData(self, imageDirs, boundingBoxFiles, meanShape, startIdx,
nImgs, mirrorFlag):
filenames = []
landmarks = []
boundingBoxes = []
for i in range(len(imageDirs)):
filenamesInDir = glob.glob(imageDirs[i] + "*.jpg")
filenamesInDir += glob.glob(imageDirs[i] + "*.png")
if boundingBoxFiles is not None:
boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'),
encoding='latin1')
for j in range(len(filenamesInDir)):
filenames.append(filenamesInDir[j])
ptsFilename = filenamesInDir[j][:-3] + "pts"
landmarks.append(utils.loadFromPts(ptsFilename))
if boundingBoxFiles is not None:
basename = path.basename(filenamesInDir[j])
boundingBoxes.append(boundingBoxDict[basename])
filenames = filenames[startIdx:startIdx + nImgs]
landmarks = landmarks[startIdx:startIdx + nImgs]
boundingBoxes = boundingBoxes[startIdx:startIdx + nImgs]
mirrorList = [False for i in range(nImgs)]
if mirrorFlag:
mirrorList = mirrorList + [True for i in range(nImgs)]
filenames = np.concatenate((filenames, filenames))
landmarks = np.vstack((landmarks, landmarks))
boundingBoxes = np.vstack((boundingBoxes, boundingBoxes))
self.origLandmarks = landmarks
self.filenames = filenames
self.mirrors = mirrorList
self.meanShape = meanShape
self.boundingBoxes = boundingBoxes
def __getitem__(
self, i
): # __getitem__ to support the indexing such that dataset[i] can be used to get ith sample
# 1. load image and kps
image = cv2.imread(self.imglists[i])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
h, w, c = image.shape
kps_path = self.imglists[i][:-4] + '.pts'
kps = loadFromPts(kps_path)
# 2. data augmentation
if self.phase == 'train':
# rotate
angle = random.randint(-self.r, self.r)
r_kps = rotatepoints(kps, [w / 2, h / 2], angle) # 逆时针旋转angle度
# norm kps to [0,res]
#旋转之后的kps的方框
left = np.min(r_kps[:, 0]) # 所有行的第0列
right = np.max(r_kps[:, 0])
top = np.min(r_kps[:, 1]) # 所有行的第1列
bot = np.max(r_kps[:, 1])
r_kps -= [left, top] # 坐标转换到[0,-]
r_kps[:, 0] *= self.res / (right - left) # 坐标转换到[0,res]
r_kps[:, 1] *= self.res / (bot - top)
# scale
s = random.uniform(0.9, 1.2) # uniform()方法将随机生成浮点数，它在 [x, y) 范围内
# make scale around center
dx = (1 - s) * self.res * 0.5 # res*0.5-s*res*0.5 缩放前的中心-缩放后的中心
s_kps = r_kps * s + [dx, dx]
# translation
dx = random.uniform(-self.res * 0.1, self.res * 0.1)
dy = random.uniform(-self.res * 0.1, self.res * 0.1)
t_kps = s_kps + [dx, dy]
# procrustes analysis 从两组关键点间分析出变换矩阵用于图像的变换
d, Z, tform = procrustes(
t_kps, kps
) # a dict specifying the rotation, translation and scaling that maps X --> Y
M = np.zeros([2, 3], dtype=np.float32)
M[:2, :2] = tform['rotation'].T * tform['scale']
M[:, 2] = tform['translation']
img = cv2.warpAffine(image, M,
(self.res, self.res)) # 仿射变换 将图像按照关键点变换
new_kps = np.dot(
kps, tform['rotation']) * tform['scale'] + tform['translation']
else:
# enlarge box
box = get_gtbox(kps)
box = enlarge_box(box, 0.05)
xmin, ymin, xmax, ymax = box
src = np.array([[xmin, ymin], [xmin, ymax], [xmax, ymin],
[xmax, ymax]])
dst = np.array([[0, 0], [0, self.res - 1], [self.res - 1, 0],
[self.res - 1, self.res - 1]])
# procrustes analysis
d, Z, tform = procrustes(dst, src)
M = np.zeros([2, 3], dtype=np.float32)
M[:2, :2] = tform['rotation'].T * tform['scale']
M[:, 2] = tform['translation']
img = cv2.warpAffine(image, M, (self.res, self.res))
new_kps = np.dot(
kps, tform['rotation']) * tform['scale'] + tform['translation']
if self.phase == 'train':
# flip
if random.random() > 0.5:
img = img[:, ::-1] # 左右翻转
new_kps = flippoints(new_kps, self.res)
# resize
if random.random() > 0.8:
new_res = int(self.res * 0.75)
img = cv2.resize(img, (new_res, new_res))
img = cv2.resize(img, (self.res, self.res))
if self.target_type == 'heatmap':
num_points = kps.shape[0]
new_kps = new_kps.astype(np.int32)
target = np.zeros([num_points, self.res, self.res])
for n in range(num_points):
target[n] = draw_gaussian(target[n],
new_kps[n],
sigma=self.gamma) # 构造训练的heatmap的标签
target = torch.from_numpy(
target).float() # 将numpy格式转换成torch.tensor格式
else:
target = torch.from_numpy(new_kps).float() # 回归landmark
# img to tensor
img = self.transform(
img.copy()
) # transforms.ToTensor() Converts a PIL Image or numpy.ndarray (H x W x C) in the range [0, 255] to
#a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0].
if self.phase == 'train':
img[0, :, :].mul_(random.uniform(0.7, 1.3)).clamp_(0, 1)
img[1, :, :].mul_(random.uniform(0.7, 1.3)).clamp_(0, 1)
img[2, :, :].mul_(random.uniform(0.7, 1.3)).clamp_(0, 1)
return img, target, torch.from_numpy(new_kps), tform
plt.figure()
plt.grid()
plt.xlim([0, 0.35])
plt.ylim([0, 1])
plt.plot(np.concatenate([[0], sorted_error]), np.arange(num_images+1, dtype=np.float)/num_images, color='darkorange', lw=2, label='cumulative curve')
plt.savefig('images/plot2.png', format='png')
'''
# print(evaluateBatchError(Ytest.reshape([-1, 68, 2]), Landmark.reshape([-1, 68, 2]), Batch_size))
for i in range(num_images):
A_temp = np.linalg.inv(A[i])
t_temp = np.dot(np.reshape(-t[i], (1, 2)), A_temp)
landmark = Landmark[i]
landmark = np.reshape(np.dot(landmark, A_temp) + t_temp, [68, 2])
ptsFilename = testSet.filenames[i][:-3] + "pts"
groundtruth = utils.loadFromPts(ptsFilename)
error = evaluateError(landmark, groundtruth)
# print(error)
errs.append(np.mean(error))
'''
img = cv2.imread(testSet.filenames[i])
for point in range(68):
cv2.circle(img,( int(landmark[point][0]), int(landmark[point][1])),1,(0,255,255),3);
cv2.imwrite(os.path.join("../data/vis",testSet.filenames[i][-14:]),img)
'''
sorted_error = np.sort(errs, axis=0)
plt.figure()
def PrepareData(self, imageDirs, boundingBoxFiles, meanShape, startIdx, nImgs, mirrorFlag):
"""获取训练集的boundingbox\imagelist\meanshape\landmark\mirrors存入相应成员变量中
Args:
imageDirs: 训练集所有图片所在路径
boudingBoxFiles: boudingBoxFile文件路径 .pkl文件
meanshape: 均值图像
startIdx: 从训练集中选取此次训练的第一个图像的id
nImgs: 此次训练的图像个数
mirrorFlag: 还不懂是干什么用的
"""
pass
print('Preparing Data......')
filenames = []
landmarks = []
boundingBoxes = []
# 获取图片名列表、lanmark标注列表、boundingbox列表分别存入以上三个变量中
for i in range(len(imageDirs)):
print('Preparing Data......%dth dataset.....'%(i))
filenamesInDir = glob.glob(imageDirs[i] + "*.jpg")
filenamesInDir += glob.glob(imageDirs[i] + "*.png")
# open() rb: 以二进制的格式打开一个文件用于只读。文件指针将会放在文件的开头。这是默认模式，返回一个file对象
# .pkl文件就是对于每一个图片创建一个dict元素,key为图片名去掉后缀，value为bounding box的四个顶点的坐标
# pickle.load() 最后返回一个dictionary
if boundingBoxFiles is not None:
boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'), encoding='iso-8859-1')
for j in range(len(filenamesInDir)):
print('Preparing Data......%dth image.....'%(j))
filenames.append(filenamesInDir[j])
# 获取标注点文件名.pts，将landmark转换为x,y坐标的array数组录入到landmarks中 68*2
ptsFilename = filenamesInDir[j][:-3] + "pts"
landmarks.append(utils.loadFromPts(ptsFilename))
if boundingBoxFiles is not None:
basename = path.basename(filenamesInDir[j])
boundingBoxes.append(boundingBoxDict[basename])
filenames = filenames[startIdx : startIdx + nImgs]
landmarks = landmarks[startIdx : startIdx + nImgs]
boundingBoxes = boundingBoxes[startIdx : startIdx + nImgs]
mirrorList = [False for i in range(nImgs)]
if mirrorFlag:
#最后得到的是一个2倍的images长度的mirrorlist,前面的mirrorlist值为false即不做镜像变换，后面的为true做镜像变换
mirrorList = mirrorList + [True for i in range(nImgs)]
# 拼接 axis default is 0，即按行拼接
filenames = np.concatenate((filenames, filenames))
# 按行拼接返回数组，作用同上
landmarks = np.vstack((landmarks, landmarks))
boundingBoxes = np.vstack((boundingBoxes, boundingBoxes))
# prepare data 进行一系列的初始化
self.origLandmarks = landmarks
self.filenames = filenames
# self.mirrors 代表是否被镜像了
self.mirrors = mirrorList
self.meanShape = meanShape
self.boundingBoxes = boundingBoxes
print('Preparing Data finished.')
def PrepareData(self, imageDirs, boundingBoxFiles, meanShape, startIdx,
nImgs, mirrorFlag):
filenames = [] #此list和filenamesInDir完全一样，可以去除
landmarks = [] #此list用于存储特征点坐标
boundingBoxes = [] #此list用于存储bbx
# import pdb; pdb.set_trace()
'''
# skip the missing pic
missing = []
missingfile = open('missing.txt')
lines = missingfile.readlines()
for line in lines:
if line[-1] == "\n":
missing.append(line[:-1])
else:
missing.append(line)
# print(len(missing))
# print(missing)
'''
for i in range(len(imageDirs)):
filenamesInDir = glob.glob(imageDirs[i] + "*.jpg")
filenamesInDir += glob.glob(imageDirs[i] + "*.png")
# import pdb; pdb.set_trace()
if boundingBoxFiles is not None:
# boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'),encoding='bytes')
boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'),
encoding='latin1')
# print(boundingBoxDict)
for j in range(len(filenamesInDir)):
# print(filenamesInDir[j])
'''
index = filenamesInDir[j].rfind("/")
picname = filenamesInDir[j][index+1:]
if picname in missing:
# print(picname)
continue
'''
filenames.append(filenamesInDir[j])
ptsFilename = filenamesInDir[j][:-3] + "pts"
landmarks.append(utils.loadFromPts(ptsFilename))
if boundingBoxFiles is not None:
basename = path.basename(filenamesInDir[j])
if boundingBoxDict.get(basename) == None:
print(basename)
continue
boundingBoxes.append(boundingBoxDict[basename])
# nImgs = len(filenames) # 在原作者代码中，nImgs=2，所以只能用两张图片，翻转+pertubation=40张图片进行训练。
# print(nImgs)
filenames = filenames[startIdx:startIdx + nImgs]
landmarks = landmarks[startIdx:startIdx + nImgs]
boundingBoxes = boundingBoxes[startIdx:startIdx + nImgs]
mirrorList = [False for i in range(nImgs)]
if mirrorFlag:
mirrorList = mirrorList + [True for i in range(nImgs)]
filenames = np.concatenate((filenames, filenames))
landmarks = np.vstack((landmarks, landmarks)) # 合并
boundingBoxes = np.vstack((boundingBoxes, boundingBoxes))
self.origLandmarks = landmarks
self.filenames = filenames
self.mirrors = mirrorList
self.meanShape = meanShape
self.boundingBoxes = boundingBoxes