def __call__(self, fname):
with Image.open(fname) as img:
img = img.convert('RGB')
img_size = Size(img.size[0], img.size[1])
if self.dataset == 'VOC':
#fname : '~/VOC/JPEGImages/002222.jpg'
with open(
os.path.join(fname[:-21], 'Annotations',
fname[-10:-4] + '.xml'), 'r') as f:
doc = lxml.etree.parse(f)
boxes = []
objects = doc.xpath('/annotation/object')
for obj in objects:
label = obj.xpath('name')[0].text
xmin = float(obj.xpath('bndbox/xmin')[0].text)
xmax = float(obj.xpath('bndbox/xmax')[0].text)
ymin = float(obj.xpath('bndbox/ymin')[0].text)
ymax = float(obj.xpath('bndbox/ymax')[0].text)
labels = self.label2idx[label]
center, size = abs2prop(
xmin, xmax, ymin, ymax,
img_size) # make gt to 0~1 important!!!!!!!!!!!!!!
box = Box(obj, labels, center, size)
boxes.append(box)
elif self.dataset == 'KITTI':
with open(
os.path.join(fname[:-18], 'label_2',
fname[-10:-4] + '.txt'), 'r') as fp:
objs = [line.split(' ') for line in fp.readlines()]
boxes = []
for obj in objs:
if not obj[0] == 'DontCare':
xmin = float(obj[4])
ymin = float(obj[5])
xmax = float(obj[6])
ymax = float(obj[7])
label = self.label2idx[obj[0]]
center, size = abs2prop(
xmin, xmax, ymin, ymax, img_size
) # Convert the absolute min-max box bound to proportional center-width bounds
# (/300 , /1300) -> (0~1, 0~1)
box = Box(obj[0], label, center, size)
boxes.append(box)
sample = Sample(fname, boxes, img_size)
return img, labels, sample
def transform_box(box, orig_size, new_size, h_off, w_off):
#---------------------------------------------------------------------------
# Compute the new coordinates of the box
#---------------------------------------------------------------------------
xmin, xmax, ymin, ymax = prop2abs(box.center, box.size, orig_size)
xmin += w_off
xmax += w_off
ymin += h_off
ymax += h_off
#---------------------------------------------------------------------------
# Check if the center falls within the image
#---------------------------------------------------------------------------
width = xmax - xmin
height = ymax - ymin
new_cx = xmin + int(width / 2)
new_cy = ymin + int(height / 2)
if new_cx < 0 or new_cx >= new_size.w:
return None
if new_cy < 0 or new_cy >= new_size.h:
return None
center, size = abs2prop(xmin, xmax, ymin, ymax, new_size)
return Box(box.label, box.labelid, center, size)
def __build_sample_list(self, root, annot_files, dataset_name):
image_root = root + '/JPEGImages/'
samples = []
for fn in tqdm(annot_files, desc=dataset_name, unit='samples'):
with open(fn, 'r') as f:
doc = lxml.etree.parse(f)
filename = image_root + doc.xpath(
'/annotation/filename')[0].text
if not os.path.exists(filename):
continue
img = cv2.imread(filename)
imgsize = Size(img.shape[1], img.shape[0])
boxes = []
objects = doc.xpath('/annotation/object')
for obj in objects:
label = obj.xpath('name')[0].text
if label == myObject:
xmin = int(float(obj.xpath('bndbox/xmin')[0].text))
xmax = int(float(obj.xpath('bndbox/xmax')[0].text))
ymin = int(float(obj.xpath('bndbox/ymin')[0].text))
ymax = int(float(obj.xpath('bndbox/ymax')[0].text))
center, size = abs2prop(xmin, xmax, ymin, ymax,
imgsize)
box = Box(label, self.lname2id[label], center, size)
boxes.append(box)
if not boxes:
continue
sample = Sample(filename, boxes, imgsize)
samples.append(sample)
return samples
def __build_sample_list(self, root, annot_files):
"""
Build a list of samples for the VOC dataset (either trainval or test)
"""
image_root = os.path.join(root, 'rgb-images')
samples = []
#-----------------------------------------------------------------------
# Process each annotated sample
#-----------------------------------------------------------------------
for fn in tqdm(annot_files, desc='ucf_24_frame', unit='samples'):
act = fn.split('/')[4]
video = fn.split('/')[5]
frame_id = fn.split('/')[-1][:-4]
image_path = os.path.join(image_root, act, video,
'{}.jpg'.format(frame_id))
#---------------------------------------------------------------
# Get the file dimensions
#---------------------------------------------------------------
if not os.path.exists(image_path):
continue
img = cv2.imread(image_path)
imgsize = Size(img.shape[1], img.shape[0])
#---------------------------------------------------------------
# Get boxes for all the objects
#---------------------------------------------------------------
boxes = []
with open(fn, 'r') as fin:
objects = fin.readlines()
for line in objects:
line = line[:-1]
#-----------------------------------------------------------
# Get the properties of the box and convert them to the
# proportional terms
#-----------------------------------------------------------
obj = line.split(' ')
label = int(obj[0]) - 1
xmin = int(float(obj[1]))
ymin = int(float(obj[2]))
xmax = int(float(obj[3]))
ymax = int(float(obj[4]))
center, size = abs2prop(xmin, xmax, ymin, ymax, imgsize)
box = Box(self.lid2name[label], label, center, size)
boxes.append(box)
if not boxes:
continue
sample = Sample(image_path, boxes, imgsize)
samples.append(sample)
return samples
def _build_sample_list(self, root, annot_files):
image_root = root + 'VOC_Jpeg/'
image_seg_root = root + 'VOC_Segmentation/'
samples = []
for fn in tqdm(annot_files, unit='samples'):
with open(fn, 'r') as f:
doc = lxml.etree.parse(f)
filename = image_root + doc.xpath(
'/annotation/filename')[0].text
with open(fn, 'r') as f1:
doc1 = lxml.etree.parse(f1)
seg_gt = image_seg_root + doc1.xpath(
'/annotation/filename')[0].text
seg_gt = seg_gt.replace('jpg', 'png')
seg_gt_to_compare = seg_gt
#---------------------------------------------------------------
# Get the file dimensions
#---------------------------------------------------------------
if not os.path.exists(filename):
continue
img = cv2.imread(filename)
img_seg_gt = cv2.imread(seg_gt)
imgsize = Size(img.shape[1], img.shape[0])
#---------------------------------------------------------------
# Get boxes for all the objects
#---------------------------------------------------------------
boxes = []
objects = doc.xpath('/annotation/object')
for obj in objects:
#-----------------------------------------------------------
# Get the properties of the box and convert them to the
# proportional terms
#-----------------------------------------------------------
label = obj.xpath('name')[0].text
xmin = int(float(obj.xpath('bndbox/xmin')[0].text))
xmax = int(float(obj.xpath('bndbox/xmax')[0].text))
ymin = int(float(obj.xpath('bndbox/ymin')[0].text))
ymax = int(float(obj.xpath('bndbox/ymax')[0].text))
center, size = abs2prop(xmin, xmax, ymin, ymax, imgsize)
box = Box(label, self.lname2id[label], center, size)
boxes.append(box)
if not boxes:
continue
sample = Sample(filename, boxes, imgsize, seg_gt,
seg_gt_to_compare)
samples.append(sample)
return samples
def __build_sample_list(self, root, dataset_name):
"""
Build a list of samples for the VOC dataset (either trainval or test)
"""
image_root = root + '/JPEGImages/'
annot_root = root + '/Annotations/'
annot_files = glob(annot_root + '/*xml')
samples = []
#-----------------------------------------------------------------------
# Process each annotated sample
#-----------------------------------------------------------------------
for fn in tqdm(annot_files, desc=dataset_name, unit='samples'):
with open(fn, 'r') as f:
doc = lxml.etree.parse(f)
filename = image_root + doc.xpath(
'/annotation/filename')[0].text
#---------------------------------------------------------------
# Get the file dimensions
#---------------------------------------------------------------
if not os.path.exists(filename):
continue
img = cv2.imread(filename)
imgsize = Size(img.shape[1], img.shape[0])
#---------------------------------------------------------------
# Get boxes for all the objects
#---------------------------------------------------------------
boxes = []
objects = doc.xpath('/annotation/object')
for obj in objects:
#-----------------------------------------------------------
# Get the properties of the box and convert them to the
# proportional terms
#-----------------------------------------------------------
label = obj.xpath('name')[0].text
xmin = int(float(obj.xpath('bndbox/xmin')[0].text))
xmax = int(float(obj.xpath('bndbox/xmax')[0].text))
ymin = int(float(obj.xpath('bndbox/ymin')[0].text))
ymax = int(float(obj.xpath('bndbox/ymax')[0].text))
center, size = abs2prop(xmin, xmax, ymin, ymax, imgsize)
box = Box(label, self.lname2id[label], center, size)
boxes.append(box)
if not boxes:
continue
sample = Sample(filename, boxes, imgsize)
samples.append(sample)
return samples
def __build_sample_list(self, root,image_dir,info_dir):
"""
Build a list of samples for the VOC dataset (either trainval or test)
"""
samples = []
file = open(root+info_dir ,'r')
All = file.read()
file.close()
lines = All.split('\n')
print(len(lines))
#-----------------------------------------------------------------------
# Process each annotated sample
#-----------------------------------------------------------------------
i=0
while i<(len(lines)-1)/10:
#---------------------------------------------------------------
# Get the file dimensions
#---------------------------------------------------------------
filename = root+image_dir+lines[i]
img = cv2.imread(filename)
imgsize = Size(img.shape[1], img.shape[0])
#---------------------------------------------------------------
# Get boxes for all the objects
#---------------------------------------------------------------
boxes = []
i+=1
num_objects = int(lines[i])
if not num_objects:
i+=2
continue
i+=1
for obj in range(num_objects):
#-----------------------------------------------------------
# Get the properties of the box and convert them to the
# proportional terms
#-----------------------------------------------------------
label = 'face'
xmin = int(lines[i+obj].split()[0])-int(lines[i+obj].split()[2])/2
xmax = int(lines[i+obj].split()[0])+int(lines[i+obj].split()[2])/2
ymin = int(lines[i+obj].split()[1])-int(lines[i+obj].split()[3])/2
ymax = int(lines[i+obj].split()[1])+int(lines[i+obj].split()[3])/2
center, size = abs2prop(xmin, xmax, ymin, ymax, imgsize)
box = Box(label, self.lname2id[label], center, size)
boxes.append(box)
i+=num_objects
sample = Sample(filename, boxes, imgsize)
samples.append(sample)
return samples
