def parser_txt_anno(video_dir, video_id, txt_anno, track_save_dir):
subfix = ".jpg"
if (len(os.listdir(track_save_dir)) == len(os.listdir(video_dir))): return
with open(txt_anno, 'r') as f:
for index, line in enumerate(f):
img_name = str(index)
img_file = os.path.join(video_dir, img_name + subfix)
#assert os.path.exists(img_file),img_file
if not os.path.exists(img_file):
continue
img = None
img = imread(img_file)
line_list = line.split(",")
bbox = [float(x) for x in line_list]
target_box = convert_bbox_format(Rectangle(*bbox), 'center-based')
crop, scale, new_sizes = get_crops(img,
target_box,
size_z=127,
size_x=255,
context_amount=0.5)
savename = osp.join(
track_save_dir,
'{}.w.{}.h.{}.jpg'.format(img_name, int(np.rint(new_sizes[0])),
int(np.rint(new_sizes[1]))))
if osp.exists(savename):
continue
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
def parser_txt_anno(video_dir, video_id, txt_anno,track_save_dir):
if not osp.exists(track_save_dir):
os.makedirs(track_save_dir)
have_croped_list = []
else:
count = count_out_of_view_frame(txt_anno)
saved_list = glob(track_save_dir+"/*.jpg")
crop_imgs_size = len(saved_list)
origin_img_size = len(glob(video_dir+"/*.jpg"))
if((crop_imgs_size+count)==origin_img_size):
print("video already croped, skip this video")
return
else:
print("crop_imgs_size: %d, origin_img_size: %d, out-of-view: %d"%(crop_imgs_size, origin_img_size, count))
#return
have_croped_list = [ i.split('.')[0]+'.jpg' for i in saved_list]
img_files = glob(video_dir+"/*.jpg")
img_files.sort()
with open(txt_anno,'r') as f:
for index, line in enumerate(tqdm(f)):
if img_files[index] in have_croped_list:
print("img %s has been croped, skip"%(img_files[index]))
continue
img = None
img = imread(img_files[index])
if isinstance(img, type(None)):
continue
line_list = line.split(",")
bbox = [int(float(x)) for x in line_list]
#skip out-of-view frames
if bbox[2]==0 or bbox[3]==0:
print("found out-of-view frame, skip this frame")
continue
#convert from 1-based to 0-based
bbox[0] = bbox[0]-1
bbox[1] = bbox[1]-1
target_box = convert_bbox_format(Rectangle(*bbox), 'center-based')
#target_box = Rectangle(*bbox)
if target_box.width<=0 or target_box.height<=0:
print("target_box error in",txt_anno, index)
continue
crop, scale,new_sizes = get_crops(img, target_box,
size_z=127, size_x=255,
context_amount=0.5)
img_id = img_files[index].split('/')[-1].split('.')[0]
savename = osp.join(track_save_dir, '{}.w.{}.h.{}.jpg'.format(img_id,int(np.rint(new_sizes[0])),int(np.rint(new_sizes[1]))))
#print(savename)
if osp.exists(savename):
continue
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
def process_split(root_dir, save_dir, split, subdir='', ):
data_dir = osp.join(root_dir, 'Data', 'VID', split)
anno_dir = osp.join(root_dir, 'Annotations', 'VID', split, subdir)
video_names = os.listdir(anno_dir)
for idx, video in enumerate(video_names):
print('{split}-{subdir} ({idx}/{total}): Processing {video}...'.format(split=split, subdir=subdir,
idx=idx, total=len(video_names),
video=video))
video_path = osp.join(anno_dir, video)
xml_files = glob(osp.join(video_path, '*.xml'))
for xml in xml_files:
tree = ET.parse(xml)
root = tree.getroot()
folder = root.find('folder').text
filename = root.find('filename').text
# Read image
img_file = osp.join(data_dir, folder, filename + '.JPEG')
img = None
# Get all object bounding boxes
bboxs = []
for object in root.iter('object'):
bbox = object.find('bndbox')
xmax = float(bbox.find('xmax').text)
xmin = float(bbox.find('xmin').text)
ymax = float(bbox.find('ymax').text)
ymin = float(bbox.find('ymin').text)
width = xmax - xmin + 1
height = ymax - ymin + 1
bboxs.append([xmin, ymin, width, height])
for idx, object in enumerate(root.iter('object')):
id = object.find('trackid').text
class_name = object.find('name').text
track_save_dir = get_track_save_directory(save_dir, 'train', subdir, video)
mkdir_p(track_save_dir)
savename = osp.join(track_save_dir, '{}.{:02d}.crop.x.jpg'.format(filename, int(id)))
if osp.isfile(savename): continue # skip existing images
if img is None:
img = imread(img_file)
# Get crop
target_box = convert_bbox_format(Rectangle(*bboxs[idx]), 'center-based')
crop, _ = get_crops(img, target_box,
size_z=127, size_x=255,
context_amount=0.5, )
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
def process_split(root_dir, save_dir, split, subdir='', ):
data_dir = osp.join(root_dir, 'Data', 'VID', split)
anno_dir = osp.join(root_dir, 'Annotations', 'VID', split, subdir)
video_names = os.listdir(anno_dir)
for idx, video in enumerate(video_names):
print('{split}-{subdir} ({idx}/{total}): Processing {video}...'.format(split=split, subdir=subdir,
idx=idx, total=len(video_names),
video=video))
video_path = osp.join(anno_dir, video)
xml_files = glob(osp.join(video_path, '*.xml'))
for xml in xml_files:
tree = ET.parse(xml)
root = tree.getroot()
folder = root.find('folder').text
filename = root.find('filename').text
# Read image
img_file = osp.join(data_dir, folder, filename + '.JPEG')
img = None
# Get all object bounding boxes
bboxs = []
for object in root.iter('object'):
bbox = object.find('bndbox')
xmax = float(bbox.find('xmax').text)
xmin = float(bbox.find('xmin').text)
ymax = float(bbox.find('ymax').text)
ymin = float(bbox.find('ymin').text)
width = xmax - xmin + 1
height = ymax - ymin + 1
bboxs.append([xmin, ymin, width, height])
for idx, object in enumerate(root.iter('object')):
id = object.find('trackid').text
class_name = object.find('name').text
track_save_dir = get_track_save_directory(save_dir, 'train', subdir, video)
mkdir_p(track_save_dir)
savename = osp.join(track_save_dir, '{}.{:02d}.crop.x.jpg'.format(filename, int(id)))
if osp.isfile(savename): continue # skip existing images
if img is None:
img = imread(img_file)
# Get crop
target_box = convert_bbox_format(Rectangle(*bboxs[idx]), 'center-based')
crop, _ = get_crops(img, target_box, size_z=127, size_x=255, context_amount=0.01)
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
def process_split(root_dir, save_dir, split):
data_dir = osp.join(root_dir, split)
video_names = os.listdir(data_dir)
video_names = [vn for vn in video_names if '.txt' not in vn]
for idx, video in enumerate(video_names):
print('{split} ({idx}/{total}): Processing {video}...'.format(split=split,
idx=idx, total=len(video_names),
video=video))
video_path = osp.join(data_dir, video)
jpg_files = glob(osp.join(video_path, '*.jpg'))
with open(osp.join(video_path, 'groundtruth.txt')) as f:
ann_content = f.readlines()
for jpg in jpg_files:
# Read image
img_file = jpg.split('/')[-1]
img = None
# Get all object bounding boxes
jpgidx = img_file.split('.')[0]
jpgidx = int(jpgidx) - 1
ann = ann_content[jpgidx]
ann = ann.strip()
bbox = ann.split(',')
bbox = [int(float(bb)) for bb in bbox] # [xmin, ymin, w, h]
track_save_dir = osp.join(save_dir, split, video)
mkdir_p(track_save_dir)
savename = osp.join(track_save_dir, '{}.crop.x.jpg'.format(img_file))
if osp.isfile(savename):
try:
im = Image.open(savename)
continue # skip existing images
except IOError:
os.remove(savename)
if img is None:
img = imread(jpg)
# Get crop
target_box = convert_bbox_format(Rectangle(*bbox), 'center-based')
crop, _ = get_crops(img, target_box,
size_z=127, size_x=255,
context_amount=0.5)
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
def track_vot_init(self, sess, first_bbox, frame):
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [frame, bbox_feed]
frame2crop_scale = self.siamese_model.initialize(sess, input_feed)
# Storing target state
self.vot_original_target_height = bbox.height
self.vot_original_target_width = bbox.width
self.vot_search_center = np.array(
[get_center(self.x_image_size),
get_center(self.x_image_size)])
self.vot_current_target_state = TargetState(
bbox=bbox,
search_pos=self.vot_search_center,
scale_idx=int(get_center(self.num_scales)))
def init(self, sess, frame, first_bbox, logdir='/tmp'):
# Get initial target bounding box and convert to center based
self.i = 0
first_bbox = Rectangle(first_bbox[0], first_bbox[1], first_bbox[2],
first_bbox[3])
bbox = convert_bbox_format(first_bbox, 'center-based')
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [
frame, bbox_feed, self.x_image_size_init, self.search_factors_init
]
frame2crop_scale, self.image_z = self.siamese_model.initialize(
sess, input_feed)
imwrite(osp.join(logdir, 'aimagez.jpg'),
cv2.cvtColor(self.image_z, cv2.COLOR_RGB2BGR))
# Storing target state
self.original_target_height = bbox.height
self.original_target_width = bbox.width
self.search_center = np.array([
get_center(self.x_image_size_init),
get_center(self.x_image_size_init)
])
self.current_target_state = TargetState(
bbox=bbox,
search_pos=self.search_center,
scale_idx=int(get_center(self.num_scales)))
self.store_thresh = 0.9
self.conf_thresh = 0.7
self.bound_thresh = 0.5
self.sup_thresh = 0.1
self.mem_count = 0
self.update_delay = 0
self.lost = 0
self.x_image_size = self.x_image_size_init
self.image_c = None
self.moved2border = False
self.prev_score = self.conf_thresh + 0.01
return True
def initialize(self, sess, first_bbox, frame, logdir='/tmp'):
"""Runs tracking on a single image sequence."""
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [frame, bbox_feed]
self.frame2crop_scale = self.siamese_model.initialize(sess, input_feed)
# Storing target state
self.original_target_height = bbox.height
self.original_target_width = bbox.width
self.search_center = np.array([get_center(self.x_image_size),
get_center(self.x_image_size)])
self.current_target_state = TargetState(bbox=bbox,
search_pos=self.search_center,
scale_idx=int(get_center(self.num_scales)))
self.logdir = logdir
self.frame_cnt = 0
def parser_xml_anno(img_file, xml_anno, track_save_dir):
tree = ET.parse(xml_anno)
root = tree.getroot()
img = None
# Get all object bounding boxes
bboxs = []
for object in root.iter('object'):
bbox = object.find('bndbox')
xmax = float(bbox.find('xmax').text)
xmin = float(bbox.find('xmin').text)
ymax = float(bbox.find('ymax').text)
ymin = float(bbox.find('ymin').text)
width = xmax - xmin + 1
height = ymax - ymin + 1
bboxs.append([xmin, ymin, width, height])
for idx, object in enumerate(root.iter('object')):
#id = object.find('trackid').text
if img is None:
img = cv2.imread(img_file)
target_box = convert_bbox_format(Rectangle(*bboxs[idx]),
'center-based')
crop, scale, new_sizes = get_crops(img,
target_box,
size_z=127,
size_x=255,
context_amount=0.5)
index_sub = "_" + str(idx) if idx > 0 else ""
save_dir = track_save_dir + index_sub
if not os.path.exists(save_dir):
os.makedirs(save_dir)
savename = os.path.join(
save_dir, '0.w.{}.h.{}.jpg'.format(int(np.rint(new_sizes[0])),
int(np.rint(new_sizes[1]))))
if osp.exists(savename):
continue
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
def track(self, sess, frame, logdir='/tmp'):
"""Runs tracking on a single image."""
i = self.i = self.i + 1
current_target_state = self.current_target_state
original_target_height = self.original_target_height
original_target_width = self.original_target_width
search_center = self.search_center
mem_count = self.mem_count
moved2border = self.moved2border
update_delay = self.update_delay + 1
lost = self.lost + 1
image_c = self.image_c
x_image_size = self.x_image_size
search_factors = self.search_factors_init
conf_thresh = self.conf_thresh
bound_thresh = self.bound_thresh
sup_thresh = self.sup_thresh
prev_score = self.prev_score
hi, wi, _ = frame.shape
h_ratio = current_target_state.bbox.height / hi
w_ratio = current_target_state.bbox.width / wi
t_i_ratio = max([h_ratio, w_ratio])
if prev_score < conf_thresh:
x_image_size += 100
#x_image_size = min(x_image_size, ((1. - t_i_ratio) * 1.8 + 1.) * self.x_image_size_init)
if t_i_ratio < 0.05:
x_image_size = min(x_image_size, 555)
elif t_i_ratio > 0.6:
x_image_size = min(x_image_size, 255)
elif t_i_ratio > 0.4:
x_image_size = min(x_image_size, 355)
else:
x_image_size = min(x_image_size, 455)
else:
x_image_size = self.x_image_size_init
num_scales = len(search_factors)
bbx = current_target_state.bbox.x
bby = current_target_state.bbox.y
bbw = current_target_state.bbox.width
bbh = current_target_state.bbox.height
bbox_feed = [bby, bbx, bbh, bbw]
if i > 1:
top = (current_target_state.bbox.y -
(current_target_state.bbox.height / 2) < 10)
left = (current_target_state.bbox.x -
(current_target_state.bbox.width / 2) < 10)
bottom = (current_target_state.bbox.y +
(current_target_state.bbox.height / 2) > hi - 10)
right = (current_target_state.bbox.x +
(current_target_state.bbox.width / 2) > wi - 10)
if top or left or bottom or right:
if not prev_score < bound_thresh:
moved2border = True
if not moved2border:
current_target_state.bbox = Rectangle(
wi / 2, hi / 2, current_target_state.bbox.width,
current_target_state.bbox.height)
bbox_feed = [
current_target_state.bbox.y,
current_target_state.bbox.x,
current_target_state.bbox.height,
current_target_state.bbox.width
]
else:
if not prev_score < bound_thresh:
moved2border = False
if t_i_ratio < 0.3 and lost > 5:
lost = 0
diffy = hi * 0.5 - bbox_feed[0]
diffx = wi * 0.5 - bbox_feed[1]
bbox_feed = [
diffy * 0.25 + bbox_feed[0], diffx * 0.25 + bbox_feed[1],
bbox_feed[2], bbox_feed[3]
]
current_target_state.bbox = Rectangle(bbox_feed[1], bbox_feed[0],
bbox_feed[3], bbox_feed[2])
input_feed = [frame, bbox_feed, x_image_size, search_factors]
outputs, metadata = self.siamese_model.inference_step(sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
re_out = np.around(1 / (1 + np.exp(-response)), 2)
if np.max(re_out) < conf_thresh:
x_image_sizeb4 = x_image_size
x_image_size += 100
#x_image_size_l = ((1. - t_i_ratio) * 1.8 + 1.) * self.x_image_size_init
if t_i_ratio < 0.05:
x_image_size_l = 555
elif t_i_ratio > 0.6:
x_image_size_l = 255
elif t_i_ratio > 0.4:
x_image_size_l = 355
else:
x_image_size_l = 455
if not x_image_size > x_image_size_l:
input_feed = [frame, bbox_feed, x_image_size, search_factors]
outputs, metadata = self.siamese_model.inference_step(
sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
re_out = np.around(1 / (1 + np.exp(-response)), 2)
else:
x_image_size = x_image_sizeb4
# Choose the scale whole response map has the highest peak
if num_scales > 1:
response_max = np.max(response * (re_out > sup_thresh),
axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones(
(num_scales))
current_scale_idx = int(get_center(num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
if max(response_penalized) == 0.:
best_scale = 1
else:
best_scale = np.argmax(response_penalized)
else:
best_scale = 0
response = response[best_scale]
re_out = re_out[best_scale]
with np.errstate(all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
response = response * (re_out > sup_thresh)
window = np.dot(np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (
1 - window_influence) * response + window_influence * self.window
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(), response.shape)
prev_score = re_out[r_max, c_max]
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(
np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config[
'embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[
best_scale]
# Position within frame in frame coordinates
y = current_target_state.bbox.y
x = current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
y = np.round(y)
x = np.round(x)
# Target scale damping and saturation
target_scale = current_target_state.bbox.height / original_target_height
search_factor = search_factors[best_scale]
scale_damp = self.track_config[
'scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 + scale_damp * search_factor)
# Some book keeping
search_center = np.array(
[get_center(x_image_size),
get_center(x_image_size)])
height = original_target_height * target_scale
width = original_target_width * target_scale
current_target_state.bbox = Rectangle(x, y, width, height)
current_target_state.scale_idx = best_scale
current_target_state.search_pos = search_center + disp_instance_input
assert 0 <= current_target_state.search_pos[0] < x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= current_target_state.search_pos[1] < x_image_size, \
'target position in feature space should be no larger than input image size'
if self.log_level > 0:
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][best_scale].astype(
np.uint8)
y_search, x_search = current_target_state.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search, target_width_search,
target_height_search)
bbox_search = convert_bbox_format(bbox_search, 'top-left-based')
xmin = bbox_search.x.astype(np.int32)
ymin = bbox_search.y.astype(np.int32)
xmax = xmin + bbox_search.width.astype(np.int32)
ymax = ymin + bbox_search.height.astype(np.int32)
cv2.rectangle(image_cropped, (xmin, ymin), (xmax, ymax),
(255, 0, 0), 2)
text = str(prev_score)
cv2.putText(image_cropped,
text, (xmin, ymin),
cv2.FONT_HERSHEY_SIMPLEX,
1.0, (255, 0, 0),
lineType=cv2.LINE_AA)
imwrite(osp.join(logdir, 'image_cropped{}.jpg'.format(i)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
if prev_score > self.store_thresh:
bbox_feed = [
current_target_state.bbox.y, current_target_state.bbox.x,
current_target_state.bbox.height,
current_target_state.bbox.width
]
self.siamese_model.update_mem(sess, [
frame, bbox_feed, self.x_image_size_init,
self.search_factors_init, mem_count
])
mem_count += 1
if mem_count > 4 or (mem_count > 0 and update_delay > 5):
self.siamese_model.update(sess)
mem_count = 0
update_delay = 0
if prev_score > bound_thresh:
lost = 0
self.mem_count = mem_count
self.update_delay = update_delay
self.moved2border = moved2border
self.lost = lost
self.x_image_size = x_image_size
self.prev_score = prev_score
self.current_target_state = current_target_state
reported_bbox = convert_bbox_format(current_target_state.bbox,
'top-left-based')
#return prev_score>0.4, reported_bbox, prev_score
return prev_score > 0.4, reported_bbox
def track_init(self, first_bbox, first_frame_image_path):
print(first_frame_image_path)
first_frame_image = safe_imread(first_frame_image_path)
self.first_frame_image = cv2.cvtColor(
first_frame_image,
cv2.COLOR_BGR2RGB) if self.image_use_rgb else first_frame_image
self.first_bbox = convert_bbox_format(
Rectangle(first_bbox[0], first_bbox[1], first_bbox[2],
first_bbox[3]), 'center-based')
first_image_crop, _, target_size = get_crops(self.first_frame_image,
self.first_bbox,
self.z_image_size,
self.x_image_size, 0.5)
cx = (self.x_image_size - 1) / 2.0
cy = (self.x_image_size - 1) / 2.0
gt_examplar_box = np.array([
cx - target_size[0] / 2.0, cy - target_size[1] / 2.0,
cx + target_size[0] / 2.0, cy + target_size[1] / 2.0
], np.float32)
self.img_height, self.img_width, _ = self.first_frame_image.shape
if self.save_video:
video_name = first_frame_image_path.split('/')[-3] + '.mp4'
fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
result_dir = os.path.join(Project_root,
self.track_config['log_dir'])
if not os.path.exists(result_dir):
os.makedirs(result_dir)
video_path = os.path.join(result_dir, video_name)
print("save video into %s" % (video_path))
self.video = cv2.VideoWriter(video_path, fourcc, 30,
(self.img_width, self.img_height))
def center_crop(img, crop_size=127):
img_shape = np.shape(img)
center_y = (img_shape[0] - 1) // 2
center_x = (img_shape[1] - 1) // 2
h = crop_size
w = crop_size
croped_img = img[center_y - h // 2:center_y + h // 2 + 1,
center_x - w // 2:center_x + w // 2 + 1]
assert (croped_img.shape[0] == crop_size)
return croped_img
self.first_image_examplar = center_crop(first_image_crop,
self.z_image_size)
shift_y = (self.x_image_size - self.z_image_size) // 2
shift_x = shift_y
x1 = gt_examplar_box[0] - shift_x
y1 = gt_examplar_box[1] - shift_y
x2 = gt_examplar_box[2] - shift_x
y2 = gt_examplar_box[3] - shift_y
self.gt_examplar_boxes = np.reshape(np.array([x1, y1, x2, y2]), [1, 4])
self.current_target_state = TargetState(bbox=self.first_bbox)
self.window = np.tile(
np.outer(np.hanning(self.score_size),
np.hanning(self.score_size)).flatten(),
5) #5 is the number of aspect ratio anchors
def track(self, first_bbox, frames, bSaveImage=False, SavePath='/tmp'):
#1. init the tracker
self.track_init(first_bbox, frames[0])
include_first = self.track_config['include_first']
# Run tracking loop
reported_bboxs = []
examplar = np.reshape(self.first_image_examplar,
[1, self.z_image_size, self.z_image_size, 3])
cost_time_dict = {
'load_img': 0.0,
'crop_img': 0.0,
'sess_run': 0.0,
'post_process': 0.0
}
for i, filename in tqdm(enumerate(frames)):
if i > 0 or include_first: # We don't really want to process the first image unless intended to do so.
load_img_start = time.time()
bgr_img = safe_imread(filename)
load_img_end = time.time()
cost_time_dict['load_img'] += load_img_end - load_img_start
crop_img_start = time.time()
current_img = cv2.cvtColor(
bgr_img,
cv2.COLOR_BGR2RGB) if self.image_use_rgb else bgr_img
instance_img, scale_x, _ = get_crops(
current_img, self.current_target_state.search_box,
self.z_image_size, self.x_image_size, 0.5)
instance = np.reshape(
instance_img, [1, self.x_image_size, self.x_image_size, 3])
crop_img_end = time.time()
cost_time_dict['crop_img'] += crop_img_end - crop_img_start
sess_run_start = time.time()
if self.model.model_config.get('BinWindow', False):
boxes, scores = self.sess.run(
[self.model.topk_bboxes, self.model.topk_scores],
feed_dict={
self.model.examplar_feed: examplar,
self.model.instance_feed: instance,
self.model.gt_examplar_boxes:
self.gt_examplar_boxes
})
else:
boxes, scores = self.sess.run(
[self.model.topk_bboxes, self.model.topk_scores],
feed_dict={
self.model.examplar_feed: examplar,
self.model.instance_feed: instance
})
sess_run_end = time.time()
cost_time_dict['sess_run'] += sess_run_end - sess_run_start
post_process_start = time.time()
def padded_size(w, h):
context = 0.5 * (w + h)
return np.sqrt((w + context) * (h + context))
#boxes: 1*NA*4 score: 1*Na
boxes = boxes[0] #NA*4
scores = scores[0] #NA*2
scales = padded_size(
(boxes[:, 2] - boxes[:, 0]) / scale_x,
(boxes[:, 3] - boxes[:, 1]) / scale_x) #Na
ratios = (boxes[:, 3] - boxes[:, 1]) / (boxes[:, 2] -
boxes[:, 0])
scale_change = scales / self.current_target_state.scale
scale_change = np.maximum(scale_change, 1.0 / scale_change)
ratio_change = ratios / (self.current_target_state.ratio)
ratio_change = np.maximum(ratio_change, 1.0 / ratio_change)
scale_penalty = np.exp(-(scale_change * ratio_change - 1) *
self.track_config['penalty_k'])
pscores = scores * scale_penalty
window_influence = self.track_config['window_influence']
wpscores = pscores * (
1 - window_influence) + self.window * window_influence
max_index = np.argmax(wpscores)
corrdinates = boxes[max_index] #Top1
#print("Tracking %d/%d with tracking score:%.2f, wpscore: %.2f"%(i+1, len(frames), scores[max_index],wpscores[max_index]))
# Position within frame in frame coordinates
res_box = Rectangle(*corrdinate_to_bbox(corrdinates))
center_x = (self.x_image_size - 1.0) / 2
center_y = center_x
delta_x = (res_box.x - center_x) / scale_x
delta_y = (res_box.y - center_y) / scale_x
w = res_box.width / scale_x
h = res_box.height / scale_x
y = self.current_target_state.target_box.y + delta_y
x = self.current_target_state.target_box.x + delta_x
#update seach bbox
alpha = self.track_config[
'search_scale_smooth_factor'] * pscores[max_index]
belta = 0.0
new_search_cx = max(
min(
self.img_width,
self.current_target_state.target_box.x * belta +
(1.0 - belta) * x), 0.0)
new_search_cy = max(
min(
self.img_height,
self.current_target_state.target_box.y * belta +
(1.0 - belta) * y), 0.0)
new_search_w = max(
10.0,
min(
self.current_target_state.target_box.width *
(1.0 - alpha) + alpha * w, self.img_width))
new_search_h = max(
10.0,
min(
self.current_target_state.target_box.height *
(1.0 - alpha) + alpha * h, self.img_height))
self.current_target_state.target_box = Rectangle(
new_search_cx, new_search_cy, new_search_w, new_search_h)
self.current_target_state.scale = padded_size(
new_search_w, new_search_h)
self.current_target_state.ratio = new_search_h * 1.0 / new_search_w
#auto increase the search region if max score is lower than the conf_threshold
if (scores[max_index] < self.conf_threshold
and self.auto_increase):
increase_w = min(new_search_w * 1.5, self.img_width)
increase_h = min(new_search_h * 1.5, self.img_height)
self.current_target_state.search_box = Rectangle(
new_search_cx, new_search_cy, increase_w, increase_h)
else:
self.current_target_state.search_box = self.current_target_state.target_box
#save and show tracking process
if bSaveImage:
cv2.imwrite(SavePath + "/" + os.path.basename(frames[i]),
bgr_img)
elif self.save_video:
x1, y1, x2, y2 = bbox_to_corrdinate(
self.current_target_state.search_box)
cv2.rectangle(bgr_img, (int(x1), int(y1)),
(int(x2), int(y2)), (0, 255, 0), 2)
cv2.putText(bgr_img, "%.2f" % (scores[max_index]),
(int(x1), int(y1)), 0, 1, (0, 255, 0), 2)
self.video.write(bgr_img)
elif self.show_video:
x1, y1, x2, y2 = bbox_to_corrdinate(
self.current_target_state.search_box)
cv2.rectangle(bgr_img, (int(x1), int(y1)),
(int(x2), int(y2)), (0, 255, 0), 2)
cv2.putText(bgr_img, "%.2f" % (scores[max_index]),
(int(x1), int(y1)), 0, 1, (0, 255, 0), 2)
cv2.imshow("Tracker", bgr_img)
cv2.waitKey(10)
else:
pass
post_process_end = time.time()
cost_time_dict[
'post_process'] += post_process_end - post_process_start
else:
x1, y1, x2, y2 = bbox_to_corrdinate(
self.current_target_state.search_box)
cv2.rectangle(self.first_frame_image, (int(x1), int(y1)),
(int(x2), int(y2)), (255, 255, 255), 2)
#cv2.imshow("Tracker",cv2.cvtColor(self.first_frame_image, cv2.COLOR_RGB2BGR))
#cv2.imshow("Target",self.first_frame_image)
#cv2.waitKey(100)
reported_bbox = convert_bbox_format(
self.current_target_state.target_box, 'top-left-based')
reported_bboxs.append(reported_bbox)
for key in cost_time_dict:
cost_time_dict[key] /= len(frames)
#print(cost_time_dict)
return reported_bboxs
def track(self, sess, frame):
bbox_feed = [
self.current_target_state.bbox.y, self.current_target_state.bbox.x,
self.current_target_state.bbox.height,
self.current_target_state.bbox.width
]
input_feed = [frame, bbox_feed]
outputs, metadata = self.siamese_model.inference_step(sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_max = np.max(response, axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones(
(self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
best_scale = np.argmax(response_penalized)
else:
best_scale = 0
response = response[best_scale]
with np.errstate(all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
if self.window is None:
window = np.dot(np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (
1 - window_influence) * response + window_influence * self.window
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(), response.shape)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(
np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config[
'embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[
best_scale]
# Position within frame in frame coordinates
y = self.current_target_state.bbox.y
x = self.current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
# Target scale damping and saturation
target_scale = self.current_target_state.bbox.height / self.original_target_height
search_factor = self.search_factors[best_scale]
scale_damp = self.track_config[
'scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 + scale_damp * search_factor)
target_scale = np.maximum(0.2, np.minimum(5.0, target_scale))
# Some book keeping
height = self.original_target_height * target_scale
width = self.original_target_width * target_scale
self.current_target_state.bbox = Rectangle(x, y, width, height)
self.current_target_state.scale_idx = best_scale
self.current_target_state.search_pos = self.search_center + disp_instance_input
assert 0 <= self.current_target_state.search_pos[0] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= self.current_target_state.search_pos[1] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
reported_bbox = convert_bbox_format(self.current_target_state.bbox,
'top-left-based')
self.frame_cnt += 1
if self.log_level > 0:
np.save(osp.join(self.logdir, 'num_frames.npy'), [self.frame_cnt])
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][best_scale].astype(
np.uint8)
# Note that imwrite in cv2 assumes the image is in BGR format.
# However, the cropped image returned by TensorFlow is RGB.
# Therefore, we convert color format using cv2.cvtColor
cv2.imwrite(
osp.join(self.logdir,
'image_cropped{}.jpg'.format(self.frame_cnt)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
cv2.imwrite(
osp.join(self.logdir,
'image_origin{}.jpg'.format(self.frame_cnt)),
cv2.cvtColor(frame, cv2.COLOR_RGB2BGR))
np.save(
osp.join(self.logdir,
'best_scale{}.npy'.format(self.frame_cnt)),
[best_scale])
np.save(
osp.join(self.logdir, 'response{}.npy'.format(self.frame_cnt)),
response)
y_search, x_search = self.current_target_state.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search, target_width_search,
target_height_search)
bbox_search = convert_bbox_format(bbox_search, 'top-left-based')
np.save(osp.join(self.logdir, 'bbox{}.npy'.format(self.frame_cnt)),
[
bbox_search.x, bbox_search.y, bbox_search.width,
bbox_search.height
])
with open(osp.join(self.logdir, 'track_rect.txt'), 'a') as f:
rect_str = '{},{},{},{}\n'.format(int(reported_bbox[0]),
int(reported_bbox[1]),
int(reported_bbox[2]),
int(reported_bbox[3]))
f.write(rect_str)
return reported_bbox
def track_vot(self, sess, frame):
bbox_feed = [
self.vot_current_target_state.bbox.y,
self.vot_current_target_state.bbox.x,
self.vot_current_target_state.bbox.height,
self.vot_current_target_state.bbox.width
]
input_feed = [frame, bbox_feed]
outputs, metadata = self.siamese_model.inference_step(sess, input_feed)
search_scale_list = outputs['scale_xs']
response_s_c5 = outputs['response_s_c5']
response_s_c4 = outputs['response_s_c4']
response_s_c3 = outputs['response_s_c3']
response_a_c5 = outputs['response_a_c5']
response_a_c4 = outputs['response_a_c4']
response_a_c3 = outputs['response_a_c3']
response_size = response_s_c5.shape[1]
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_a_c5_max = np.max(response_a_c5)
response_a_c4_max = np.max(response_a_c4)
response_a_c3_max = np.max(response_a_c3)
response_a_c5 = response_a_c5 / response_a_c5_max
response_a_c4 = response_a_c4 / response_a_c4_max
response_a_c3 = response_a_c3 / response_a_c3_max
response_s_all = 0.7 * response_s_c5 + 0.3 * response_s_c4 + 0.1 * response_s_c3
response_a_all = 0.3 * response_a_c5 + 0.6 * response_a_c4 + 0.1 * response_a_c3
response_s_all_max = np.max(response_s_all)
response_s_all = response_s_all / response_s_all_max
response_a_all_max = np.max(response_a_all)
response_a_all = response_a_all / response_a_all_max
response = 0.3 * response_s_all + 0.7 * response_a_all
response_max = np.max(response, axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones(
(self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
best_scale = np.argmax(response_penalized)
else:
## TODO combine siamfc and alexnet
best_scale = 0
response = response[best_scale]
with np.errstate(all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
if self.window is None:
window = np.dot(np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (
1 - window_influence) * response + window_influence * self.window
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(), response.shape)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(
np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * 8 #self.model_config['embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[
best_scale]
# Position within frame in frame coordinates
y = self.vot_current_target_state.bbox.y
x = self.vot_current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
# Target scale damping and saturation
target_scale = self.vot_current_target_state.bbox.height / self.vot_original_target_height
search_factor = self.search_factors[best_scale]
scale_damp = self.track_config[
'scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 + scale_damp * search_factor)
target_scale = np.maximum(0.2, np.minimum(5.0, target_scale))
# Some book keeping
height = self.vot_original_target_height * target_scale
width = self.vot_original_target_width * target_scale
self.vot_current_target_state.bbox = Rectangle(x, y, width, height)
self.vot_current_target_state.scale_idx = best_scale
self.vot_current_target_state.search_pos = self.vot_search_center + disp_instance_input
reported_bbox = convert_bbox_format(self.vot_current_target_state.bbox,
'top-left-based')
return reported_bbox
def track(self, sess, first_bbox, frames, logdir='/tmp'):
"""Runs tracking on a single image sequence."""
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [
frames[0], bbox_feed, self.x_image_size, self.search_factors
]
frame2crop_scale, image_z = self.siamese_model.initialize(
sess, input_feed)
imwrite(osp.join(logdir, 'aimagez.jpg'),
cv2.cvtColor(image_z, cv2.COLOR_RGB2BGR))
# Storing target state
original_target_height = bbox.height
original_target_width = bbox.width
search_center = np.array(
[get_center(self.x_image_size),
get_center(self.x_image_size)])
current_target_state = TargetState(bbox=bbox,
search_pos=search_center,
scale_idx=int(
get_center(self.num_scales)))
include_first = get(self.track_config, 'include_first', False)
logging.info('Tracking include first -- {}'.format(include_first))
# Run tracking loop
reported_bboxs = []
image_c = None
x_image_size = self.x_image_size
lost = 0
moved2border = False
conf_thresh = 0.2 # 0.2
bound_thresh = 0.2 # 0.2
sup_thresh = 0.15 # 0.15
prev_score = conf_thresh + 0.01
upsample_factor = self.track_config['upsample_factor']
search_factors = self.search_factors
for i, filename in enumerate(frames):
if i > 0 or include_first:
bbox_feed = [
current_target_state.bbox.y, current_target_state.bbox.x,
current_target_state.bbox.height,
current_target_state.bbox.width
]
if prev_score > bound_thresh:
lost = 0
else:
lost += 1
if prev_score > 0.9:
self.siamese_model.update(sess, [
frames[i - 1], bbox_feed, self.x_image_size,
search_factors
])
with open(filename, 'rb') as f:
wi, hi = GetWidthAndHeight(f)
t_i_ratio = max([
current_target_state.bbox.height / hi,
current_target_state.bbox.width / wi
])
if prev_score < conf_thresh:
x_image_size += 100
#x_image_size = min(x_image_size, ((1. - t_i_ratio) * 1.6 + 1.) * self.x_image_size_init)
if t_i_ratio < 0.05:
x_image_size = min(x_image_size, 555)
elif t_i_ratio < 0.25:
x_image_size = min(x_image_size, 455)
elif t_i_ratio > 0.5:
x_image_size = min(x_image_size, 255)
else:
x_image_size = min(x_image_size, 355)
else:
x_image_size = self.x_image_size
if i > 1:
top = (current_target_state.bbox.y -
(current_target_state.bbox.height / 2) < 10)
left = (current_target_state.bbox.x -
(current_target_state.bbox.width / 2) < 10)
bottom = (current_target_state.bbox.y +
(current_target_state.bbox.height / 2) > hi - 10)
right = (current_target_state.bbox.x +
(current_target_state.bbox.width / 2) > wi - 10)
bound_flag = top or left or bottom or right
#if top or left or bottom or right:
#if not prev_score < bound_thresh:
#moved2border = True
#if not moved2border:
#current_target_state.bbox = Rectangle(wi / 2, hi / 2,
#current_target_state.bbox.width,
#current_target_state.bbox.height)
#bbox_feed = [current_target_state.bbox.y, current_target_state.bbox.x,
#current_target_state.bbox.height, current_target_state.bbox.width]
#else:
#if not prev_score < bound_thresh:
#moved2border = False
if lost > 5 and bound_flag:
lost = 0
diffy = hi * 0.5 - bbox_feed[0]
diffx = wi * 0.5 - bbox_feed[1]
bbox_feed = [
diffy * 0.25 + bbox_feed[0],
diffx * 0.25 + bbox_feed[1], bbox_feed[2], bbox_feed[3]
]
current_target_state.bbox = Rectangle(bbox_feed[1],
bbox_feed[0],
bbox_feed[3],
bbox_feed[2])
input_feed = [
filename, bbox_feed, x_image_size, search_factors
]
outputs, metadata = self.siamese_model.inference_step(
sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
re_out = np.around(1 / (1 + np.exp(-response)), 2)
if np.max(re_out) < conf_thresh and not t_i_ratio > 0.5:
x_image_sizeb4 = x_image_size
x_image_size += 100
#x_image_size_l = ((1. - t_i_ratio) * 1.6 + 1.) * self.x_image_size_init
if t_i_ratio < 0.05:
x_image_size_l = 555
elif t_i_ratio < 0.25:
x_image_size_l = 455
elif t_i_ratio > 0.5:
x_image_size_l = 255
else:
x_image_size_l = 355
if not x_image_size > x_image_size_l:
input_feed = [
filename, bbox_feed, x_image_size, search_factors
]
outputs, metadata = self.siamese_model.inference_step(
sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
re_out = np.around(1 / (1 + np.exp(-response)), 2)
else:
x_image_size = x_image_sizeb4
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_max = np.max(response * (re_out > sup_thresh),
axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones(
(self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
if max(response_penalized) == 0.:
best_scale = 1
else:
best_scale = np.argmax(response_penalized)
else:
best_scale = 0
response = response[best_scale]
re_out = re_out[best_scale]
with np.errstate(
all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
response = response * (re_out > sup_thresh)
window = np.dot(np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (1 - window_influence
) * response + window_influence * self.window
if np.max(re_out) < sup_thresh:
r_max, c_max = response.shape
r_max, c_max = int(r_max / 2), int(c_max / 2)
disp_instance_input = [0, 0]
disp_instance_frame = [0, 0]
else:
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(),
response.shape)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(
np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config[
'embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[
best_scale]
# Position within frame in frame coordinates
y = current_target_state.bbox.y
x = current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
y = np.round(y)
x = np.round(x)
prev_score = re_out[r_max, c_max]
# Target scale damping and saturation
target_scale = current_target_state.bbox.height / original_target_height
search_factor = search_factors[best_scale]
scale_damp = self.track_config[
'scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 +
scale_damp * search_factor)
# Some book keeping
search_center = np.array(
[get_center(x_image_size),
get_center(x_image_size)])
height = original_target_height * target_scale
width = original_target_width * target_scale
current_target_state.bbox = Rectangle(x, y, width, height)
current_target_state.scale_idx = best_scale
current_target_state.search_pos = search_center + disp_instance_input
assert 0 <= current_target_state.search_pos[0] < x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= current_target_state.search_pos[1] < x_image_size, \
'target position in feature space should be no larger than input image size'
if self.log_level > 0:
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][
best_scale].astype(np.uint8)
y_search, x_search = current_target_state.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search,
target_width_search,
target_height_search)
bbox_search = convert_bbox_format(bbox_search,
'top-left-based')
# Add score colormap
image_cropped = outputs['image_cropped'][
best_scale].astype(np.uint8)
#im_shape = image_cropped.shape
#re_shape = response_size / upsample_factor * self.model_config['embed_config']['stride']
#pad = int((im_shape[0] - re_shape) / 2)
#response_crop = imresize(re_out, [im_shape[0]-2*pad, im_shape[1]-2*pad])
#response_crop = np.pad(response_crop, ((pad, pad), (pad, pad)), 'constant')
#response_crop = response_crop / response_crop.max()
#response_crop = np.uint8(response_crop * 255)
#cmap = cv2.cvtColor(cv2.applyColorMap(response_crop, cv2.COLORMAP_JET), cv2.COLOR_BGR2RGB)
#image_cropped = cv2.addWeighted(cmap, 0.3, image_cropped, 0.5, 0)
xmin = bbox_search.x.astype(np.int32)
ymin = bbox_search.y.astype(np.int32)
xmax = xmin + bbox_search.width.astype(np.int32)
ymax = ymin + bbox_search.height.astype(np.int32)
cv2.rectangle(image_cropped, (xmin, ymin), (xmax, ymax),
(255, 0, 0), 2)
text = str(prev_score)
cv2.putText(image_cropped,
text, (xmin, ymin),
cv2.FONT_HERSHEY_SIMPLEX,
1.0, (255, 0, 0),
lineType=cv2.LINE_AA)
imwrite(osp.join(logdir, 'image_cropped{}.jpg'.format(i)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
#if image_c is not None:
#his_dir = logdir + '_his'
#if not osp.exists(his_dir):
#os.mkdir(his_dir)
#image_c_p = np.concatenate([np.expand_dims(image_z, 0)] + image_c, 2)[0]
#image_c_p = np.uint8(image_c_p)
#imwrite(osp.join(his_dir, 'image{}.jpg'.format(i)),
#cv2.cvtColor(image_c_p, cv2.COLOR_RGB2BGR))
reported_bbox = convert_bbox_format(current_target_state.bbox,
'top-left-based')
reported_bboxs.append(reported_bbox)
return reported_bboxs
def process_split(root_dir, save_dir, split):
data_dir = osp.join(root_dir, split)
video_names = os.listdir(data_dir)
video_names = [vn for vn in video_names if '.txt' not in vn]
for idx, video in enumerate(video_names):
print('{split} ({idx}/{total}): Processing {video}...'.format(
split=split, idx=idx, total=len(video_names), video=video))
video_path = osp.join(data_dir, video)
jpg_files = glob(osp.join(video_path, '*.jpg'))
with open(osp.join(video_path, 'groundtruth.txt')) as f:
ann_content = f.readlines()
track_save_dir = osp.join(save_dir, split, video)
mkdir_p(track_save_dir)
fw = open(osp.join(track_save_dir, 'groundtruth.txt'), 'w')
copyfile(osp.join(video_path, 'absence.label'),
osp.join(track_save_dir, 'absence.label'))
copyfile(osp.join(video_path, 'cover.label'),
osp.join(track_save_dir, 'cover.label'))
copyfile(osp.join(video_path, 'cut_by_image.label'),
osp.join(track_save_dir, 'cut_by_image.label'))
copyfile(osp.join(video_path, 'meta_info.ini'),
osp.join(track_save_dir, 'meta_info.ini'))
for i, jpg in enumerate(jpg_files):
# Read image
img_file = jpg.split('/')[-1]
img = None
# Get all object bounding boxes
jpgidx = img_file.split('.')[0]
jpgidx = int(jpgidx) - 1
ann = ann_content[jpgidx]
ann = ann.strip()
bbox = ann.split(',')
bbox = [int(float(bb)) for bb in bbox] # [xmin, ymin, w, h]
## bbox ####
annk = ann_content[i]
annk = annk.strip()
bboxk = annk.split(',')
bboxk = [int(float(bb)) for bb in bboxk] # [xmin, ymin, w, h]
w = bboxk[2]
h = bboxk[3]
context_amount = 0.5
size_z = 127
size_x = 271
wc_z = w + context_amount * (w + h)
hc_z = h + context_amount * (w + h)
s_z = np.sqrt(wc_z * hc_z)
scale_z = size_z / s_z
d_search = (size_x - size_z) / 2
pad = d_search / scale_z
s_x = s_z + 2 * pad
wn = int(w * size_x / s_x)
hn = int(h * size_x / s_x)
#if wn < 1 or hn < 1:
#if wn == 0:
#wn = 1
#if hn == 0:
#hn = 1
#ratio = wn / hn
#if ratio > 1.:
#newbb = [int(135-wn/2), int(135-hn/2), 85, int(85. / ratio)]
#else:
#newbb = [int(135-wn/2), int(135-hn/2), int(85. * ratio), 85]
#else:
#newbb = [int(135-wn/2), int(135-hn/2), wn, hn]
if wn < 1:
wn = 1
if hn < 1:
hn = 1
newbb = [int(135 - wn / 2), int(135 - hn / 2), wn, hn]
fw.write(','.join(str(e) + '.0000' for e in newbb) + '\n')
## bbox ####
savename = osp.join(track_save_dir, '{}.jpg'.format(img_file))
if osp.isfile(savename):
try:
im = Image.open(savename)
continue # skip existing images
except IOError:
os.remove(savename)
if img is None:
img = imread(jpg)
# Get crop
target_box = convert_bbox_format(Rectangle(*bbox), 'center-based')
crop, _ = get_crops(img,
target_box,
size_z=127,
size_x=271,
context_amount=0.5)
imwrite(savename, crop, [int(cv2.IMWRITE_JPEG_QUALITY), 90])
fw.close()
def main(_):
# load model
model_config, _, track_config = load_cfgs(CHECKPOINT)
track_config["log_level"] = 0
track_config["is_video"] = True
g = tf.Graph()
with g.as_default():
model = inference_wrapper.InferenceWrapper()
restore_fn = model.build_graph_from_config(model_config, track_config,
CHECKPOINT)
g.finalize()
if not os.path.isdir(track_config['log_dir']):
tf.logging.info('Creating inference directory: %s',
track_config['log_dir'])
mkdir_p(track_config['log_dir'])
gpu_options = tf.GPUOptions(allow_growth=True)
sess_config = tf.ConfigProto(gpu_options=gpu_options)
with tf.Session(graph=g, config=sess_config) as sess:
restore_fn(sess)
tracker = Tracker(model,
model_config=model_config,
track_config=track_config)
video_name = os.path.basename(FLAGS.video_path)
video_log_dir = os.path.join(track_config["log_dir"], video_name)
mkdir_p(video_log_dir)
if str(FLAGS.video_path) in ["0", "1"]:
# read from camera
video_path = int(FLAGS.video_path)
with_camera = True
else:
# read from video
video_path = glob(os.path.join(FLAGS.video_path, "*.mp4"))[0]
with_camera = False
video_capture = cv2.VideoCapture(video_path)
bb = [-1, -1, -1, -1]
cv2.namedWindow("template")
cv2.setMouseCallback("template", draw_init_box, bb)
trajectory = []
f_count = 0
f_rate = 0
start_time = time.time()
while True:
# capture frame by frame
ret_, frame = video_capture.read()
if ret_ == False:
continue
f_width, f_height = [
int(a) for a in FLAGS.video_resolution.split("*")
]
try:
o_frame = cv2.resize(frame, (f_width, f_height),
interpolation=cv2.INTER_CUBIC)
except:
break
i_frame = cv2.cvtColor(o_frame, cv2.COLOR_BGR2RGB)
# cv2.imwrite("test.jpg",o_frame)
# pdb.set_trace()
if f_count == 0: # initialize the tracker
# wait for drawing init box
while True:
init_frame = o_frame.copy()
cv2.imshow("template", init_frame)
k = cv2.waitKey(0)
if k == 32: # space
cx = int((bb[0] + bb[2]) / 2)
cy = int((bb[1] + bb[3]) / 2)
w = int(bb[2] - bb[0])
h = int(bb[3] - bb[1])
# Rectangle: [x,y,width,height]
init_bb = Rectangle(cx - 1, cy - 1, w,
h) # 0-index in python
draw_box(init_frame, init_bb, "exemplar")
break
first_box = convert_bbox_format(init_bb, "center-based")
bbox_feed = [
first_box.y, first_box.x, first_box.height, first_box.width
]
input_feed = [i_frame, bbox_feed]
frame2crop_scale = tracker.siamese_model.initialize(
sess, input_feed)
# Storing target state
original_target_height = first_box.height
original_target_width = first_box.width
search_center = np.array([
get_center(tracker.x_image_size),
get_center(tracker.x_image_size)
])
current_target_state = TargetState(
bbox=first_box,
search_pos=search_center,
scale_idx=int(get_center(tracker.num_scales)))
# setup initialized params
current_param = {
"original_target_width": original_target_width,
"original_target_height": original_target_height,
"search_center": search_center,
"current_target_state": current_target_state
}
bbox, current_param = tracker.track_frame(sess, i_frame,
current_param,
video_log_dir)
# add overlays
end_time = time.time()
f_rate = int(1 / (end_time - start_time))
start_time = time.time()
draw_box(o_frame, bbox)
cv2.putText(o_frame,
str(f_rate) + "fps", (10, 30),
cv2.FONT_HERSHEY_SIMPLEX,
1, (0, 0, 255),
thickness=2,
lineType=2)
trajectory.append(bbox)
f_count += 1
cv2.imshow("Real-time Ouput", o_frame)
cv2.imshow("template", init_frame)
# if f_count > 30:
# cv2.imwrite("test.jpg",o_frame)
# pdb.set_trace()
if cv2.waitKey(1) & 0xFF == ord("q"):
cv2.imwrite("./assets/instance.jpg", o_frame)
cv2.imwrite("./assets/exemplar.jpg", init_frame)
break
video_capture.release()
cv2.destroyAllWindows()
# save track results
# pdb.set_trace()
with open(os.path.join(video_log_dir, "track_rect.txt"), "w") as f:
for region in trajectory:
rect_str = "{},{},{},{}\n".format(region.x + 1, region.y + 1,
region.width, region.height)
f.write(rect_str)
def track(self, sess, first_bbox, frames, logdir='/tmp'):
"""Runs tracking on a single image sequence."""
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [frames[0], bbox_feed]
frame2crop_scale = self.siamese_model.initialize(sess, input_feed)
# Storing target state
original_target_height = bbox.height
original_target_width = bbox.width
search_center = np.array([get_center(self.x_image_size),
get_center(self.x_image_size)])
current_target_state = TargetState(bbox=bbox,
search_pos=search_center,
scale_idx=int(get_center(self.num_scales)))
include_first = get(self.track_config, 'include_first', False)
logging.info('Tracking include first -- {}'.format(include_first))
# Run tracking loop
reported_bboxs = []
for i, filename in enumerate(frames):
if i > 0 or include_first: # We don't really want to process the first image unless intended to do so.
bbox_feed = [current_target_state.bbox.y, current_target_state.bbox.x,
current_target_state.bbox.height, current_target_state.bbox.width]
input_feed = [filename, bbox_feed]
outputs, metadata = self.siamese_model.inference_step(sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_max = np.max(response, axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones((self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
best_scale = np.argmax(response_penalized)
else:
best_scale = 0
response = response[best_scale]
with np.errstate(all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
if self.window is None:
window = np.dot(np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (1 - window_influence) * response + window_influence * self.window
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(),
response.shape)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config['embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[best_scale]
# Position within frame in frame coordinates
y = current_target_state.bbox.y
x = current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
# Target scale damping and saturation
target_scale = current_target_state.bbox.height / original_target_height
search_factor = self.search_factors[best_scale]
scale_damp = self.track_config['scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 + scale_damp * search_factor)
target_scale = np.maximum(0.2, np.minimum(5.0, target_scale))
# Some book keeping
height = original_target_height * target_scale
width = original_target_width * target_scale
current_target_state.bbox = Rectangle(x, y, width, height)
current_target_state.scale_idx = best_scale
current_target_state.search_pos = search_center + disp_instance_input
assert 0 <= current_target_state.search_pos[0] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= current_target_state.search_pos[1] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
if self.log_level > 0:
np.save(osp.join(logdir, 'num_frames.npy'), [i + 1])
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][best_scale].astype(np.uint8)
# Note that imwrite in cv2 assumes the image is in BGR format.
# However, the cropped image returned by TensorFlow is RGB.
# Therefore, we convert color format using cv2.cvtColor
imwrite(osp.join(logdir, 'image_cropped{}.jpg'.format(i)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
np.save(osp.join(logdir, 'best_scale{}.npy'.format(i)), [best_scale])
np.save(osp.join(logdir, 'response{}.npy'.format(i)), response)
y_search, x_search = current_target_state.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search, target_width_search, target_height_search)
bbox_search = convert_bbox_format(bbox_search, 'top-left-based')
np.save(osp.join(logdir, 'bbox{}.npy'.format(i)),
[bbox_search.x, bbox_search.y, bbox_search.width, bbox_search.height])
reported_bbox = convert_bbox_format(current_target_state.bbox, 'top-left-based')
reported_bboxs.append(reported_bbox)
return reported_bboxs
def track(self, sess, first_bbox, frames, logdir='/tmp'):
"""Runs tracking on a single image sequence."""
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
smooth_rate = self.track_config['smooth']
update_interval = self.track_config['update_interval']
feature_balance = self.track_config['feature_balance']
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [frames[0], bbox_feed]
frame2crop_scale = self.siamese_model.initialize(sess, input_feed)
examplar = self.siamese_model.get_examplar(sess, input_feed)
examplar_smooth = examplar
st_template = []
for i in range(self.siamese_model.train_config['time_range']):
st_template.append(examplar)
st_template_np = np.array(st_template)
self.siamese_model.update_st_template_step(sess, st_template_np)
# Storing target state
original_target_height = bbox.height
original_target_width = bbox.width
search_center = np.array(
[get_center(self.x_image_size),
get_center(self.x_image_size)])
current_target_state = TargetState(bbox=bbox,
search_pos=search_center,
scale_idx=int(
get_center(self.num_scales)))
include_first = get(self.track_config, 'include_first', False)
logging.info('Tracking include first -- {}'.format(include_first))
# Set padding for refining search region
img = mpimg.imread(frames[0])
context_amount = self.track_config['context_amount']
size_z = self.model_config['z_image_size']
size_x = self.track_config['x_image_size']
padding_h = 10
padding_w = 10
if original_target_height / original_target_width > 2: #2
padding_h = 1.4 #1.4
padding_w = 6
# Run tracking loop
reported_bboxs = []
for i, filename in enumerate(frames):
if i > 0 or include_first: # We don't really want to process the first image unless intended to do so.
bbox_feed = [
current_target_state.bbox.y, current_target_state.bbox.x,
current_target_state.bbox.height,
current_target_state.bbox.width
]
input_feed = [filename, bbox_feed]
outputs, metadata = self.siamese_model.inference_step(
sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response2 = outputs['response2']
response_size = response.shape[1]
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_max = np.max(response2, axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones(
(self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
best_scale = np.argmax(response_penalized)
else:
best_scale = 0
response = response[best_scale]
response2 = response2[best_scale]
response = feature_balance * response + (
1 - feature_balance) * response2
with np.errstate(
all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
if self.window is None:
window = np.dot(
np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (1 - window_influence
) * response + window_influence * self.window
# Refine the response
base_z_size = np.array([
current_target_state.bbox.height,
current_target_state.bbox.width
])
base_z_context_size = base_z_size + context_amount * np.sum(
base_z_size)
base_s_z = np.sqrt(
np.prod(base_z_context_size)) # Canonical size
base_scale_z = size_z / base_s_z
d_search = (size_x - size_z) / 2.0
base_pad = d_search / base_scale_z
base_s_x = base_s_z + 2 * base_pad
if base_s_x / current_target_state.bbox.height > padding_h:
start_h = np.ceil(
response_size *
(base_s_x -
current_target_state.bbox.height * padding_h) /
(2 * base_s_x))
end_h = np.floor(response_size - start_h)
start_h = np.int(start_h)
end_h = np.int(end_h)
response[0:start_h, :] = 0
response[end_h:-1, :] = 0
if base_s_x / current_target_state.bbox.width > padding_w:
start_w = np.ceil(
response_size *
(base_s_x -
current_target_state.bbox.width * padding_w) /
(2 * base_s_x))
end_w = np.floor(response_size - start_w)
start_w = np.int(start_w)
end_w = np.int(end_w)
response[:, :start_w] = 0
response[:, end_w:] = 0
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(),
response.shape)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(
np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config[
'embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[
best_scale]
# Position within frame in frame coordinates
y = current_target_state.bbox.y
x = current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
# Target scale damping and saturation
target_scale = current_target_state.bbox.height / original_target_height
search_factor = self.search_factors[best_scale]
scale_damp = self.track_config[
'scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 +
scale_damp * search_factor)
target_scale = np.maximum(0.2, np.minimum(5.0, target_scale))
# Some book keeping
height = original_target_height * target_scale
width = original_target_width * target_scale
current_target_state.bbox = Rectangle(x, y, width, height)
current_target_state.scale_idx = best_scale
current_target_state.search_pos = search_center + disp_instance_input
# Update the spatial-temporal template using gcn
if i % update_interval == 0:
bbox_feed = [
current_target_state.bbox.y,
current_target_state.bbox.x,
current_target_state.bbox.height,
current_target_state.bbox.width
]
input_feed = [filename, bbox_feed]
current_examplar = self.siamese_model.get_examplar(
sess, input_feed)
# examplar_smooth[2:4,2:4,:] = current_examplar[2:4,2:4,:]
examplar_smooth = current_examplar
current_examplar = smooth_rate * examplar_smooth + (
1 - smooth_rate) * examplar
st_template.pop(1)
st_template.append(current_examplar)
st_template_np = np.array(st_template)
self.siamese_model.update_st_template_step(
sess, st_template_np)
assert 0 <= current_target_state.search_pos[0] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= current_target_state.search_pos[1] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
if self.log_level > 0:
np.save(osp.join(logdir, 'num_frames.npy'), [i + 1])
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][
best_scale].astype(np.uint8)
# Note that imwrite in cv2 assumes the image is in BGR format.
# However, the cropped image returned by TensorFlow is RGB.
# Therefore, we convert color format using cv2.cvtColor
imwrite(osp.join(logdir, 'image_cropped{}.jpg'.format(i)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
np.save(osp.join(logdir, 'best_scale{}.npy'.format(i)),
[best_scale])
np.save(osp.join(logdir, 'response{}.npy'.format(i)),
response)
y_search, x_search = current_target_state.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search,
target_width_search,
target_height_search)
bbox_search = convert_bbox_format(bbox_search,
'top-left-based')
np.save(osp.join(logdir, 'bbox{}.npy'.format(i)), [
bbox_search.x, bbox_search.y, bbox_search.width,
bbox_search.height
])
reported_bbox = convert_bbox_format(current_target_state.bbox,
'top-left-based')
reported_bboxs.append(reported_bbox)
return reported_bboxs
def track(self, sess, first_bbox, frames, logdir='/tmp'):
"""Runs tracking on a single image sequence."""
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [frames[0], bbox_feed]
frame2crop_scale = self.siamese_model.initialize(sess, input_feed)
# Storing target state
original_target_height = bbox.height
original_target_width = bbox.width
search_center = np.array(
[get_center(self.x_image_size),
get_center(self.x_image_size)])
current_target_state = TargetState(bbox=bbox,
search_pos=search_center,
scale_idx=int(
get_center(self.num_scales)))
include_first = get(self.track_config, 'include_first', False)
logging.info('Tracking include first -- {}'.format(include_first))
# Run tracking loop
reported_bboxs = []
for i, filename in enumerate(frames):
if i > 0 or include_first: # We don't really want to process the first image unless intended to do so.
bbox_feed = [
current_target_state.bbox.y, current_target_state.bbox.x,
current_target_state.bbox.height,
current_target_state.bbox.width
]
input_feed = [filename, bbox_feed]
outputs, metadata = self.siamese_model.inference_step(
sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_max = np.max(response, axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones(
(self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
best_scale = np.argmax(response_penalized)
if np.max(response_max) < 0:
logging.warning('MAX_RESPONSE LESS THAN ZERO!')
# best_scale = current_scale_idx
else:
best_scale = 0
response = response[best_scale]
with np.errstate(
all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
if self.window is None:
window = np.dot(
np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (1 - window_influence
) * response + window_influence * self.window
# Find maximum response
r_max, c_max = np.unravel_index(response.argmax(),
response.shape)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(
np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config[
'embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[
best_scale]
# Position within frame in frame coordinates
y = current_target_state.bbox.y
x = current_target_state.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
# Target scale damping and saturation
target_scale = current_target_state.bbox.height / original_target_height
search_factor = self.search_factors[best_scale]
scale_damp = self.track_config[
'scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 +
scale_damp * search_factor)
target_scale = np.maximum(0.2, np.minimum(5.0, target_scale))
# Some book keeping
height = original_target_height * target_scale
width = original_target_width * target_scale
current_target_state.bbox = Rectangle(x, y, width, height)
current_target_state.scale_idx = best_scale
current_target_state.search_pos = search_center + disp_instance_input
assert 0 <= current_target_state.search_pos[0] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= current_target_state.search_pos[1] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
if self.log_level > 0:
np.save(osp.join(logdir, 'num_frames.npy'), [i + 1])
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][
best_scale].astype(np.uint8)
# Note that imwrite in cv2 assumes the image is in BGR format.
# However, the cropped image returned by TensorFlow is RGB.
# Therefore, we convert color format using cv2.cvtColor
imwrite(osp.join(logdir, 'image_cropped{}.jpg'.format(i)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
np.save(osp.join(logdir, 'best_scale{}.npy'.format(i)),
[best_scale])
np.save(osp.join(logdir, 'response{}.npy'.format(i)),
response)
y_search, x_search = current_target_state.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search,
target_width_search,
target_height_search)
bbox_search = convert_bbox_format(bbox_search,
'top-left-based')
np.save(osp.join(logdir, 'bbox{}.npy'.format(i)), [
bbox_search.x, bbox_search.y, bbox_search.width,
bbox_search.height
])
reported_bbox = convert_bbox_format(current_target_state.bbox,
'top-left-based')
reported_bboxs.append(reported_bbox)
return reported_bboxs
def track(self, sess, first_bbox, frames, logdir='/tmp'):
"""Runs tracking on a single image sequence."""
# Get initial target bounding box and convert to center based
bbox = convert_bbox_format(first_bbox, 'center-based')
print(frames)
# Feed in the first frame image to set initial state.
bbox_feed = [bbox.y, bbox.x, bbox.height, bbox.width]
input_feed = [frames[0], bbox_feed]
frame2crop_scale = self.siamese_model.initialize(sess, input_feed)
# Storing target state
original_target_height = bbox.height
original_target_width = bbox.width
search_center = np.array([get_center(self.x_image_size),
get_center(self.x_image_size)])
current_target_state = TargetState(bbox=bbox,
search_pos=search_center,
scale_idx=int(get_center(self.num_scales)))
include_first = get(self.track_config, 'include_first', False)
logging.info('Tracking include first -- {}'.format(include_first))
# Run tracking loop
reported_bboxs = []
output_json={} #dump all bboxes in this output file
for i, filename in enumerate(frames):
if i > 0 or include_first: # We don't really want to process the first image unless intended to do so.
bbox_feed = [current_target_state.bbox.y, current_target_state.bbox.x,
current_target_state.bbox.height, current_target_state.bbox.width]
input_feed = [filename, bbox_feed]
outputs, metadata = self.siamese_model.inference_step(sess, input_feed)
search_scale_list = outputs['scale_xs']
response = outputs['response']
response_size = response.shape[1]
# Choose the scale whole response map has the highest peak
if self.num_scales > 1:
response_max = np.max(response, axis=(1, 2))
penalties = self.track_config['scale_penalty'] * np.ones((self.num_scales))
current_scale_idx = int(get_center(self.num_scales))
penalties[current_scale_idx] = 1.0
response_penalized = response_max * penalties
best_scale = np.argmax(response_penalized)
else:
best_scale = 0
response = response[best_scale]
#print(response)
with np.errstate(all='raise'): # Raise error if something goes wrong
response = response - np.min(response)
response = response / np.sum(response)
if self.window is None:
window = np.dot(np.expand_dims(np.hanning(response_size), 1),
np.expand_dims(np.hanning(response_size), 0))
self.window = window / np.sum(window) # normalize window
window_influence = self.track_config['window_influence']
response = (1 - window_influence) * response + window_influence * self.window
# Find maximum response
srtd=response.argsort(axis=None)
v = response.argmax()
r_max, c_max = np.unravel_index(v,
response.shape)
if not osp.exists(osp.join(logdir,"Intermediate")):
os.mkdir(osp.join(logdir,"Intermediate"))
to_save = np.interp(response,(response.min(),response.max()),(0,255))
cv2.imwrite(osp.join(logdir,"Intermediate",f"response_{i}.png"),to_save)
to_save = to_save.reshape(to_save.shape[0],to_save.shape[1],1)
ret,thresh1 = cv2.threshold(to_save,185,255,cv2.THRESH_BINARY)
cv2.imwrite(osp.join(logdir,"Intermediate",f"response_{i}_thresh.png"),thresh1)
image = np.uint8(thresh1.copy())
cnts = cv2.findContours(image, cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)
cnts = imutils.grab_contours(cnts)
backtorgb = cv2.cvtColor(image,cv2.COLOR_GRAY2RGB)
image = cv2.drawContours(backtorgb, cnts, -1, (0, 255, 0), 2)
cv2.imwrite(osp.join(logdir,"Intermediate",f"response_{i}_cntrs.png"),image)
centres=[]
for c in cnts:
M = cv2.moments(c)
cX = int(M["m10"] / M["m00"])
cY = int(M["m01"] / M["m00"])
centres.append((cY,cX,False))
centres.append((r_max,c_max,True))
#print(centres)
#cts_copy = copy(current_target_state)
#cts_copy2 = copy(current_target_state)
output_json[filename]=[]
for (r_max,c_max,to_deep_copy) in centres:
if to_deep_copy:
cts_copy = deepcopy(current_target_state)
else:
cts_copy = copy(current_target_state)
# Convert from crop-relative coordinates to frame coordinates
p_coor = np.array([r_max, c_max])
# displacement from the center in instance final representation ...
disp_instance_final = p_coor - get_center(response_size)
# ... in instance feature space ...
upsample_factor = self.track_config['upsample_factor']
disp_instance_feat = disp_instance_final / upsample_factor
# ... Avoid empty position ...
r_radius = int(response_size / upsample_factor / 2)
disp_instance_feat = np.maximum(np.minimum(disp_instance_feat, r_radius), -r_radius)
# ... in instance input ...
disp_instance_input = disp_instance_feat * self.model_config['embed_config']['stride']
# ... in instance original crop (in frame coordinates)
disp_instance_frame = disp_instance_input / search_scale_list[best_scale]
# Position within frame in frame coordinates
y = cts_copy.bbox.y
x = cts_copy.bbox.x
y += disp_instance_frame[0]
x += disp_instance_frame[1]
# Target scale damping and saturation
target_scale = cts_copy.bbox.height / original_target_height
search_factor = self.search_factors[best_scale]
scale_damp = self.track_config['scale_damp'] # damping factor for scale update
target_scale *= ((1 - scale_damp) * 1.0 + scale_damp * search_factor)
target_scale = np.maximum(0.2, np.minimum(5.0, target_scale))
# Some book keeping
height = original_target_height * target_scale
width = original_target_width * target_scale
cts_copy.bbox = Rectangle(x, y, width, height)
cts_copy.scale_idx = best_scale
cts_copy.search_pos = search_center + disp_instance_input
assert 0 <= cts_copy.search_pos[0] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
assert 0 <= cts_copy.search_pos[1] < self.x_image_size, \
'target position in feature space should be no larger than input image size'
if self.log_level > 0 and to_deep_copy:
np.save(osp.join(logdir, 'num_frames.npy'), [i + 1])
# Select the image with the highest score scale and convert it to uint8
image_cropped = outputs['image_cropped'][best_scale].astype(np.uint8)
# Note that imwrite in cv2 assumes the image is in BGR format.
# However, the cropped image returned by TensorFlow is RGB.
# Therefore, we convert color format using cv2.cvtColor
imwrite(osp.join(logdir, 'image_cropped{}.jpg'.format(i)),
cv2.cvtColor(image_cropped, cv2.COLOR_RGB2BGR))
np.save(osp.join(logdir, 'best_scale{}.npy'.format(i)), [best_scale])
np.save(osp.join(logdir, 'response{}.npy'.format(i)), response)
y_search, x_search = cts_copy.search_pos
search_scale = search_scale_list[best_scale]
target_height_search = height * search_scale
target_width_search = width * search_scale
bbox_search = Rectangle(x_search, y_search, target_width_search, target_height_search)
bbox_search = convert_bbox_format(bbox_search, 'top-left-based')
np.save(osp.join(logdir, 'bbox{}.npy'.format(i)),
[bbox_search.x, bbox_search.y, bbox_search.width, bbox_search.height])
reported_bbox = convert_bbox_format(cts_copy.bbox, 'top-left-based')
#print(f"reported bbox {reported_bbox}")
if to_deep_copy:
reported_bboxs.append(reported_bbox)
else:
rect_str = '{},{},{},{}\n'.format(reported_bbox.x + 1, reported_bbox.y + 1,
reported_bbox.width, reported_bbox.height)
arr = output_json[filename]
arr.append(rect_str)
with open(osp.join(logdir,'bboxes.json'),'w') as f:
json.dump(output_json,f,indent=4)
return reported_bboxs
