def main(label_type):
wildcard = '/*/*/' if label_type == 'train' else '/'
# dataset_path = 'data/ILSVRC2015/'
# dataset_path = '/media/yueshen/Sea_Gate!/imagenet/detection/ILSVRC2014_DET_train/'
# dataset_path = '/media/yueshen/Sea_Gate!/imagenet/detection/ILSVRC2014_DET_train/'
# dataset path on google
dataset_path = '/home/detection/ILSVRC/'
annotationPath = dataset_path + 'Annotations/'
imagePath = dataset_path + 'Data/'
if not os.path.exists(os.path.join('labels', label_type)):
os.makedirs(os.path.join('labels', label_type))
imageNameFile = open('labels/' + label_type + '/image_names.txt', 'w')
labels = []
labels = glob.glob(annotationPath + 'DET/' + label_type + wildcard +
'*.xml')
print('len labels = ', len(labels))
# # test
# labels = glob.glob(annotationPath + 'n01443537/' + '*.xml')
labels.sort()
print('label = ', len(labels))
images = [
label.replace('Annotations', 'Data').replace('xml', 'JPEG')
for label in labels
]
print(len(images))
bboxes = []
for ii, imageName in enumerate(images):
if ii % 100 == 0:
print('iter %d of %d = %.2f%%' %
(ii, len(images), ii * 1.0 / len(images) * 100))
if not DEBUG:
imageNameFile.write(imageName + '\n')
imOn = ii
label = labels[imOn]
labelTree = ET.parse(label)
# try:
imgSize = get_image_size(images[imOn])
# except:
# continue
area_cutoff = imgSize[0] * imgSize[1] * 0.01
if DEBUG:
print('\nimage name\n\n%s\n' % images[imOn])
image = cv2.imread(images[imOn])
print('image size', image.shape)
print(label)
print(labelTree)
print(labelTree.findall('object'))
for obj in labelTree.findall('object'):
bbox = obj.find('bndbox')
bbox = [
int(bbox.find('xmin').text),
int(bbox.find('ymin').text),
int(bbox.find('xmax').text),
int(bbox.find('ymax').text), imOn
]
if (bbox[3] - bbox[1]) * (bbox[2] - bbox[0]) < area_cutoff:
continue
if DEBUG:
print('name', obj.find('name').text, '\n')
print(bbox)
image = image.squeeze()
if len(image.shape) < 3:
image = np.tile(image[:, :, np.newaxis], (1, 1, 3))
drawing.drawRect(image, bbox[:-1], 3, [0, 0, 255])
bboxes.append(bbox)
if DEBUG:
if len(image.shape) == 2:
image = np.tile(image[:, :, np.newaxis], (1, 1, 3))
cv2.imshow('image', image)
cv2.waitKey(0)
bboxes = np.array(bboxes)
if not DEBUG:
np.save('labels/' + label_type + '/labels.npy', bboxes)
def draw_state(self):
from utils import drawing
scale = 8
if self.board is None:
locs = self.graph.points * scale
self.board = np.zeros(
((self.graph.yMax - self.graph.yMin) * scale,
(self.graph.xMax - self.graph.xMin) * scale),
dtype=np.uint8)
locs -= np.array([self.graph.xMin, self.graph.yMin]) * scale
for loc in locs:
drawing.drawRect(self.board, [loc[0], loc[1], loc[0], loc[1]],
scale / 2, 4)
if type(self.end_point) == list:
for end_point in self.end_point:
goal_loc = (
np.array(end_point) * np.array([scale, scale, 90]) -
np.array([self.graph.xMin, self.graph.yMin, 0]) *
scale).astype(int)
drawing.drawRect(
self.board,
[goal_loc[0], goal_loc[1], goal_loc[0], goal_loc[1]],
scale / 2, 5)
else:
goal_loc = (
np.array(self.end_point) * np.array([scale, scale, 90]) -
np.array([self.graph.xMin, self.graph.yMin, 0]) *
scale).astype(int)
goal_arrow = [
goal_loc[0] + scale / 2 * (goal_loc[2] == 90) - scale / 2 *
(goal_loc[2] == 270), goal_loc[1] + scale / 2 *
(goal_loc[2] == 0) - scale / 2 * (goal_loc[2] == 180)
]
drawing.drawRect(
self.board,
[goal_loc[0], goal_loc[1], goal_loc[0], goal_loc[1]],
scale / 2, 5)
drawing.drawRect(self.board, [
goal_arrow[0], goal_arrow[1], goal_arrow[0], goal_arrow[1]
], scale / 4, 6)
self.board[np.logical_or(self.board == 2, self.board == 3)] = 4
curr_point = np.array(self.pose[:3])
curr_loc = (curr_point * np.array([scale, scale, 90]) -
np.array([self.graph.xMin, self.graph.yMin, 0]) *
scale).astype(int)
curr_arrow = [
curr_loc[0] + scale / 2 * (curr_loc[2] == 90) - scale / 2 *
(curr_loc[2] == 270), curr_loc[1] + scale / 2 *
(curr_loc[2] == 0) - scale / 2 * (curr_loc[2] == 180)
]
drawing.drawRect(self.board,
[curr_loc[0], curr_loc[1], curr_loc[0], curr_loc[1]],
scale / 2, 2)
drawing.drawRect(
self.board,
[curr_arrow[0], curr_arrow[1], curr_arrow[0], curr_arrow[1]],
scale / 4, 3)
self.board[0, 0] = 6
return np.flipud(self.board)
Images[num_seq, ...] = tImage.copy()
Labels[num_seq, ...] = xyxyLabels.copy()
num_seq += 1
print('current seq # = ', num_seq)
# np.save('Images.npy', Images)
# np.save('Labels.npy', Labels)
print('final seq idx = ', dataset.seq_idx)
print('done!')
print('Checking images... ')
path = './test/'
idx = 15 # random
image = Images[idx, ...].copy()
labels = Labels[idx, ...].copy()
for i in range(image.shape[0]):
# print(i)
im = image[i, ...].transpose(1, 2, 0).copy()
bbox = 227 * labels[i // 2, ...].copy() / 10
print('bbox = ', bbox)
patch = drawing.drawRect(im, bbox, 1, (255, 255, 0))
# print(im.shape)
cv2.imwrite(path + str(i) + '.png', patch)
# Images_load = np.load('Images.npy')
# Labels_load = np.load('Labels.npy')
# print(Images_load.shape, Labels_load.shape)
# print('images load = ', np.sum(Images_load))
# print('label load = ', Labels_load[0,0:5,:])
def get_data_sequence(self):
try:
# Preallocate the space for the datas and labels.
dataImage = np.zeros((self.num_unrolls, self.memory_size, IMG_SIZE, IMG_SIZE, 3), dtype=np.float32)
dataMotion = np.zeros((self.num_unrolls, self.memory_size, 4), dtype=np.float32)
labelImage = np.zeros((self.num_unrolls, IMG_SIZE, IMG_SIZE, 3), dtype=np.float32)
labelMotion = np.zeros((self.num_unrolls, 4), dtype=np.float32)
# Read a new data sequence from batch cache and get the ground truth.
(batchKey, images) = self.getData()
# key = [data_idx, video_idx, track_id, image_seq]
gtKey = batchKey
initImageIndex = self.key_lookup[gtKey]
if self.debug:
print('Inital gtKey: ', gtKey)
print('')
# initial data
# key = [data_idx, video_idx, track_id, image_seq]
newKey = list(gtKey)
newKey = tuple(newKey)
imageIndex = self.key_lookup[newKey]
bbox = self.datasets[newKey[0]][imageIndex, :4].copy()
x1, y1, x2, y2 = bbox
# image
image = images[0]
object = image[int(y1):int(y2), int(x1):int(x2)]
object = cv2.resize(object, (IMG_SIZE, IMG_SIZE)) / 255.
dataImage[:] = object
# motion
height, width, _ = image.shape
"""
cx = float(int((x1 + x2) / 2) / width)
cy = float(int((y1 + y2) / 2) / height)
w = float((x2 - x1) / width)
h = float((y2 - y1) / height)
dataMotion[:] = [cx, cy, w, h]
"""
x1 = float(x1 /width)
y1 = float(y1 / height)
x2 = float(x2 / width)
y2 = float(y2 / height)
dataMotion[:] = [x1, y1, x2, y2]
# data
for unroll in range(self.num_unrolls):
if self.debug:
print('Unroll : ', unroll)
print('')
debug_image = np.zeros((self.memory_size, IMG_SIZE, IMG_SIZE, 3), dtype=np.float32)
debug_motion = np.zeros((self.memory_size, 4), dtype=np.float32)
debug_image_label = np.zeros((IMG_SIZE, IMG_SIZE, 3), dtype=np.float32)
debug_motion_label = np.zeros((4), dtype=np.float32)
for memory_idx in range(self.memory_size):
if unroll >= memory_idx:
# key = [data_idx, video_idx, track_id, image_seq]
newKey = list(gtKey)
newKey[3] += unroll - memory_idx
newKey = tuple(newKey)
imageIndex = self.key_lookup[newKey]
bbox = self.datasets[newKey[0]][imageIndex, :4].copy()
x1, y1, x2, y2 = bbox
# image
image = images[unroll - memory_idx]
object = image[int(y1):int(y2), int(x1):int(x2)]
object = cv2.resize(object, (IMG_SIZE, IMG_SIZE)) / 255.
dataImage[unroll, memory_idx] = object
# motion
height, width, _ = image.shape
"""
cx = float(int((x1 + x2) / 2) / width)
cy = float(int((y1 + y2) / 2) / height)
w = float((x2 - x1) / width)
h = float((y2 - y1) / height)
dataMotion[:] = [cx, cy, w, h]
"""
x1 = float(x1 / width)
y1 = float(y1 / height)
x2 = float(x2 / width)
y2 = float(y2 / height)
dataMotion[:] = [x1, y1, x2, y2]
if self.debug:
# debug each
debug_image1 = image.copy()
drawing.drawRect(debug_image1, [x1, y1, x2, y2], 2, [255, 0, 0])
#cv2.imshow('debug bbox', debug_image1)
path = self.image_paths[newKey[0]][newKey[-1]]
print('Memory idx : ', memory_idx)
print('gtKey : ', newKey)
print('bbox : ', bbox)
print('bbox(float) : ', [x1, y1, x2, y2])
print('Image idx : ', imageIndex)
print('Image path : ', path)
print('')
debug_image[memory_idx] = object
debug_motion[memory_idx] = [x1, y1, x2, y2]
#cv2.waitKey(0)
# label
# key = [data_idx, video_idx, track_id, image_seq]
newKey = list(gtKey)
newKey[3] += unroll + 1
newKey = tuple(newKey)
imageIndex = self.key_lookup[newKey]
bbox = self.datasets[newKey[0]][imageIndex, :4].copy()
x1, y1, x2, y2 = bbox
# image
image = images[unroll + 1]
object = image[int(y1):int(y2), int(x1):int(x2)]
object = cv2.resize(object, (IMG_SIZE, IMG_SIZE)) / 255.
labelImage[unroll] = object
# motion
height, width, _ = image.shape
"""
cx = float(int((x1 + x2) / 2) / width)
cy = float(int((y1 + y2) / 2) / height)
w = float((x2 - x1) / width)
h = float((y2 - y1) / height)
dataMotion[:] = [cx, cy, w, h]
"""
x1 = float(x1 / width)
y1 = float(y1 / height)
x2 = float(x2 / width)
y2 = float(y2 / height)
labelMotion[:] = [x1, y1, x2, y2]
if self.debug:
# debug each
#debug_image1 = image.copy()
#drawing.drawRect(debug_image1, [x1, y1, x2, y2], 2, [255, 0, 0])
#cv2.imshow('debug bbox', debug_image1)
path = self.image_paths[newKey[0]][newKey[-1]]
print('[label]')
print('gtKey : ', newKey)
print('bbox : ', bbox)
print('bbox(float) : ', [x1, y1, x2, y2])
print('Image idx : ', imageIndex)
print('Image path : ', path)
print('')
debug_image_label = object
debug_motion_label = [x1, y1, x2, y2]
plots = []
for idx in range(self.memory_size):
#print('Memory idx :', idx, debug_motion[idx])
plots.append(dataImage[unroll, idx])
subplot = np.zeros((IMG_SIZE * self.memory_size, IMG_SIZE, 3), dtype=np.float32)
cv2.vconcat(tuple(plots), subplot)
cv2.imshow('external memory', subplot)
cv2.imshow('prediction', labelImage[unroll])
cv2.waitKey(0)
dataImage = dataImage.reshape(([self.num_unrolls * self.memory_size] + list(dataImage.shape[2:])))
dataMotion = dataMotion.reshape(([self.num_unrolls * self.memory_size] + list(dataMotion.shape[2:])))
return (dataImage, dataMotion), (labelImage, labelMotion)
except Exception as e:
import traceback
traceback.print_exc()
import pdb
pdb.set_trace()
print('exception')
def main(label_type):
folder = [
'/ILSVRC2015_VID_train_0001/*/', '/ILSVRC2015_VID_train_0002/*/',
'/ILSVRC2015_VID_train_0003/*/'
]
for fol in folder:
wildcard = fol if label_type == 'train' else '/*/'
# dataset_path = 'data/ILSVRC2015/'
# dataset_path = '/media/yueshen/Sea_Gate!/imagenet/ILSVRC/'
# google root directory
dataset_path = '/home/ILSVRC/'
annotationPath = dataset_path + 'Annotations/'
imagePath = dataset_path + 'Data/'
if not DEBUG:
if not os.path.exists(os.path.join('labels', label_type)):
os.makedirs(os.path.join('labels', label_type))
imageNameFile = open(
'labels/' + label_type + '/image_names' + str(fol[-4]) +
'.txt', 'w')
videos = sorted(
glob.glob(annotationPath + 'VID/' + label_type + wildcard))
bboxes = []
imNum = 0
totalImages = len(
glob.glob(annotationPath + 'VID/' + label_type + wildcard +
'*.xml'))
print('totalImages', totalImages)
classes = {
'n01674464': 1,
'n01662784': 2,
'n02342885': 3,
'n04468005': 4,
'n02509815': 5,
'n02084071': 6,
'n01503061': 7,
'n02324045': 8,
'n02402425': 9,
'n02834778': 10,
'n02419796': 11,
'n02374451': 12,
'n04530566': 13,
'n02118333': 14,
'n02958343': 15,
'n02510455': 16,
'n03790512': 17,
'n02391049': 18,
'n02121808': 19,
'n01726692': 20,
'n02062744': 21,
'n02503517': 22,
'n02691156': 23,
'n02129165': 24,
'n02129604': 25,
'n02355227': 26,
'n02484322': 27,
'n02411705': 28,
'n02924116': 29,
'n02131653': 30,
}
for vv, video in enumerate(videos):
labels = sorted(glob.glob(video + '*.xml'))
images = [
label.replace('Annotations', 'Data').replace('xml', 'JPEG')
for label in labels
]
trackColor = dict()
for ii, imageName in enumerate(images):
if imNum % 100 == 0:
print('imNum %d of %d = %.2f%%' %
(imNum, totalImages, imNum * 100.0 / totalImages))
if not DEBUG:
# Leave off initial bit of path so we can just add parent dir to path later.
imageNameFile.write(imageName + '\n')
label = labels[ii]
labelTree = ET.parse(label)
imgSize = get_image_size(images[ii])
area = imgSize[0] * imgSize[1]
if DEBUG:
print('\n%s' % images[ii])
image = cv2.imread(images[ii])
print('video', vv, 'image', ii)
for obj in labelTree.findall('object'):
cls = obj.find('name').text
assert cls in classes
classInd = classes[cls]
occl = int(obj.find('occluded').text)
trackId = int(obj.find('trackid').text)
bbox = obj.find('bndbox')
bbox = [
int(bbox.find('xmin').text),
int(bbox.find('ymin').text),
int(bbox.find('xmax').text),
int(bbox.find('ymax').text), vv, trackId, imNum,
classInd, occl
]
if DEBUG:
print('name', obj.find('name').text, '\n')
print(bbox)
if trackId not in trackColor:
trackColor[trackId] = [
random.random() * 255 for _ in range(3)
]
drawing.drawRect(image, bbox[:4], 3,
trackColor[trackId])
bboxes.append(bbox)
if DEBUG:
cv2.imshow('image', image)
cv2.waitKey(1)
imNum += 1
bboxes = np.array(bboxes)
# Reorder by video_id, then track_id, then video image number so all labels for a single track are next to each other.
# This only matters if a single image could have multiple tracks.
order = np.lexsort((bboxes[:, 6], bboxes[:, 5], bboxes[:, 4]))
bboxes = bboxes[order, :]
if not DEBUG:
np.save('labels/' + label_type + '/labels' + str(fol[-4]) + '.npy',
bboxes)