def main():
# Avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
# --- Parse arguments from JSON file ---
hp, evaluation, run, env, design = parse_arguments()
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
image, templates_z, scores = siam.build_tracking_graph(
final_score_sz, design, env)
# --- Start Streaming from Video ---
cap = cv2.VideoCapture(env.root_sequences + '/' + sys.argv[1] + '.mp4')
ret, frame = cap.read()
if (not ret):
print "Error opening video sequence"
# --- Save Video (Optional) ---
if run.save_video:
vid_write = cv2.VideoWriter(
env.root_sequences + '/' + sys.argv[1] + '_out.avi',
cv2.VideoWriter_fourcc(*'MJPG'), 25,
(frame.shape[1], frame.shape[0]), True)
# --- Define Initial Bounding Box ---
BB = click_and_crop(frame, design.window_name)
cv2.namedWindow(design.window_name)
cv2.startWindowThread()
cv2.setMouseCallback(design.window_name, BB.callback)
cv2.imshow(design.window_name, frame)
cv2.waitKey(0)
while True:
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# ----- Define Initial Bounding Box Params & Template -----
pos_x = int((BB.refPt[0][0] + BB.refPt[1][0]) / 2) # Template Center
pos_y = int((BB.refPt[0][1] + BB.refPt[1][1]) / 2) # Template Center
target_w = int(abs(BB.refPt[1][0] - BB.refPt[0][0])) # Template Width / 2
target_h = int(abs(BB.refPt[1][1] - BB.refPt[0][1])) # Template Height / 2
# ----- Beging Tracking -----
tracker(hp, run, design, pos_x, pos_y, target_w, target_h, final_score_sz,
templates_z, scores, cap, vid_write, frame)
cap.release()
cv2.destroyAllWindows()
if run.save_video:
vid_write.release()
def evaluate(): # avoid printing TF debugging information os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # TODO: allow parameters from command line or leave everything in json files? hp, evaluation, run, env, design = parse_arguments() # Set size for use with tf.image.resize_images with align_corners=True. # For example, # [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1) # instead of # [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3) final_score_sz = hp.response_up * (design.score_sz - 1) + 1 # build TF graph once for all gt, frame_name_list, _, _ = _init_video(env, evaluation, evaluation.video) frame_sz = [i for i in cv2.imread(frame_name_list[0]).shape] siamNet = siam.Siamese(batch_size = 1); image, z_crops, x_crops, templates_z, scores, loss, _, distance_to_gt, summary, templates_x, max_pos_x, max_pos_y = siamNet.build_tracking_graph_train(final_score_sz, design, env, hp, frame_sz) pos_x, pos_y, target_w, target_h = region_to_bbox(gt[evaluation.start_frame]) bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y, target_w, target_h, final_score_sz, image, templates_z, scores, evaluation.start_frame, path_ckpt = os.path.join(design.saver_folder, design.path_ckpt), siamNet = siamNet) _, precision, precision_auc, iou = _compile_results(gt, bboxes, evaluation.dist_threshold) print(evaluation.video + \ ' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\ ' -- Precision AUC: ' + "%.2f" % precision_auc + \ ' -- IOU: ' + "%.2f" % iou + \ ' -- Speed: ' + "%.2f" % speed + ' --')
def run_SiamFCpytorch(seq, rp, bSaveImage):
hp, evaluation, run, env, design = parse_arguments()
#final_score_sz = hp.response_up * (design.score_sz - 1) + 1
final_score_sz = 265
siam = SiameseNet(env.root_pretrained, design.net)
load_net(NET_PATH, siam)
siam.cuda()
frame_name_list = seq.s_frames
init_rect = seq.init_rect
x, y, width, height = init_rect # OTB format
init_bb = Rectangle(x - 1, y - 1, float(width), float(height))
init_bb = convert_bbox_format(init_bb, 'center-based')
bboxes, speed = tracker(hp, run, design, frame_name_list, init_bb.x,
init_bb.y, init_bb.width, init_bb.height,
final_score_sz, siam, evaluation.start_frame)
trajectory = [
Rectangle(val[0] + 1, val[1] + 1, val[2], val[3]) for val in bboxes
]
result = dict()
result['res'] = trajectory
result['type'] = 'rect'
result['fps'] = speed
return result
def main(process, queue, box, video):
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build TF graph once for all
image, templates_z, scores = siam.build_tracking_graph(final_score_sz, design, env)
# read radio
# width = 640
# height = 480
# process1 = (
# ffmpeg
# .input('tcp://192.168.1.155:8300',vcodec='h264',r = 24,probesize=32,fflags="nobuffer",flags="low_delay",analyzeduration=1)
# .output('pipe:', format='rawvideo',pix_fmt="rgb24")
# .run_async(pipe_stdout=True)
# )
## model
# model_path = './frozen_inference_graph.pb'
# odapi = DetectorAPI(path_to_ckpt=model_path)
# while True :
# in_bytes = process1.stdout.read(width * height * 3)
# if not in_bytes :
# print ("none")
# video = (np.frombuffer(in_bytes, np.uint8).reshape([height, width, 3]))
# video = cv2.cvtColor(video, cv2.COLOR_RGB2BGR)
# read target from mat
# box = odapi.processFrame(video)
box[2] -= box[0]
box[3] -= box[1]
box[0] += box[2]/2
box[1] += box[3]/2
print ('box', box)
pos_x, pos_y, target_w, target_h = box[0], box[1], box[2], box[3]
tracker(hp, run, design, video, pos_x, pos_y, target_w, target_h, final_score_sz,
image, templates_z, scores, process, queue)
print ('done')
def main():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build TF graph once for all
filename, image, templates_z, scores = siam.build_tracking_graph(final_score_sz, design, env)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv):
images_arr, gt, frame_name_list, frame_sz, n_frames = _init_video(env, evaluation, videos_list[i])
starts = np.rint(np.linspace(0, n_frames - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq):
start_frame = int(starts[j])
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_, pos_x, pos_y,
target_w, target_h, final_score_sz, filename,
image, templates_z, scores, start_frame)
lengths[idx], precisions[idx], precisions_auc[idx], ious[idx] = compile_results(gt, bboxes, evaluation.dist_threshold)
print(str(i) + ' -- ' + videos_list[i] + ' -- Precision: ' + "%.2f" % precisions[idx] + ' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + ' -- IOU: ' + "%.2f" % ious[idx] + ' -- Speed: ' + "%.2f" % speed[idx] + ' --')
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_precision_auc = np.sum(precisions_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print('-- Overall stats (averaged per frame) on ' + str(nv) + ' videos (' + str(tot_frames) + ' frames) --')
print(' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision + ' -- Precisions AUC: ' + "%.2f" % mean_precision_auc + ' -- IOU: ' + "%.2f" % mean_iou + ' -- Speed: ' + "%.2f" % mean_speed + ' --')
else:
images_arr, gt, frame_name_list, _, _ = _init_video(env, evaluation, evaluation.video)
pos_x, pos_y, target_w, target_h = region_to_bbox(gt[evaluation.start_frame])
bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y, target_w, target_h, final_score_sz,
filename, image, templates_z, scores, evaluation.start_frame)
num_frames = np.size(frame_name_list)
bboxes_final = np.zeros((num_frames,4))
lk_params = dict( winSize = (5,5),
maxLevel = 2,
criteria = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 1, 0.03))
bboxes_final = np.zeros((num_frames, 4))
for i in range(1,len(images_arr)-1):
# Create some random colors
color = np.random.randint(0,255,(100,3))
# Take first frame and find corners in it
#ret, old_frame = cap.read()
frame = images_arr[i+1]
old_frame = images_arr[i]
old_gray = cv2.cvtColor(old_frame, cv2.COLOR_BGR2GRAY)
p0 = np.zeros((1,1,2), dtype=np.float32)
bbox_i = bboxes[i]
c, r, w, h = int(bbox_i[0]), int(bbox_i[1]), int(bbox_i[2]), int(bbox_i[3])
p0[0,0,0] = c
p0[0,0,1] = r
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
p1, st, err = cv2.calcOpticalFlowPyrLK(old_gray, frame_gray, p0, None, **lk_params)
good_new = p1[st==1]
bboxes_final[i,:] = p1[0][0][0], p1[0][0][1], w, h
good_old = p0[st==1]
for i,(new,old) in enumerate(zip(good_new,good_old)):
a,b = new.ravel()
c,d = old.ravel()
fig = plt.figure(1)
ax = fig.add_subplot(111)
r1 = patches.ConnectionPatch((a,b),(c,d),'data','data',arrowstyle="-|>")
r2 = patches.Circle((a,b),5,color='r')#((x,y), w, h, linewidth=2, edgecolor='r', fill=False)
ax.imshow(np.uint8(frame))
ax.add_patch(r2)
ax.add_patch(r1)
plt.ion()
plt.show()
plt.pause(0.001)
plt.clf()
old_gray = frame_gray.copy()
p0 = good_new.reshape(-1,1,2)
_, precision, precision_auc, iou = _compile_results(gt, bboxes_final, evaluation.dist_threshold)
print(evaluation.video + ' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision + ' -- Precision AUC: ' + "%.2f" % precision_auc + ' -- IOU: ' + "%.2f" % iou + ' -- Speed: ' + "%.2f" % speed + ' --')
def main():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# Load hyperparameter (hp), evaluation, run, environment (env) and design parameters from
# parameters/ directory.
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build TF graph once for all
filename, image, templates_z, scores = siam.build_tracking_graph(
final_score_sz, design, env)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv):
gt, frame_name_list, frame_sz, n_frames = _init_video(
env, evaluation, videos_list[i])
starts = np.rint(
np.linspace(0, n_frames - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq):
start_frame = int(starts[j])
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_,
pos_x, pos_y, target_w, target_h,
final_score_sz, filename, image,
templates_z, scores, start_frame)
lengths[idx], precisions[idx], precisions_auc[idx], ious[
idx] = _compile_results(gt_, bboxes,
evaluation.dist_threshold)
print str(i) + ' -- ' + videos_list[i] + \
' -- Precision: ' + "%.2f" % precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
' -- Speed: ' + "%.2f" % speed[idx] + ' --'
print
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_precision_auc = np.sum(precisions_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print '-- Overall stats (averaged per frame) on ' + str(
nv) + ' videos (' + str(tot_frames) + ' frames) --'
print ' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" % mean_iou +\
' -- Speed: ' + "%.2f" % mean_speed + ' --'
print
else:
gt, frame_name_list, _, _ = _init_video(env, evaluation,
evaluation.video)
pos_x, pos_y, target_w, target_h = region_to_bbox(
gt[evaluation.start_frame])
bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y,
target_w, target_h, final_score_sz, filename,
image, templates_z, scores,
evaluation.start_frame)
_, precision, precision_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
print evaluation.video + \
' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + \
' -- Speed: ' + "%.2f" % speed + ' --'
print
def main():
hp, evaluation, run, env, design = parse_arguments()
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
siam = SiameseNet(env.root_pretrained, design.net)
if torch.cuda.is_available():
siam = siam.cuda()
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
video_list = [v for v in os.listdir(dataset_folder) if not v[0] == '.']
video_list.sort()
nv = np.size(video_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precision_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv):
print('video: %d' % (i + 1))
gt, frame_name_list, frame_sz, n_frame = _init_video(
env, evaluation, video_list[i])
starts = np.rint(
np.linspace(0, n_frame - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq):
starts_frame = int(starts[j])
gt_ = gt[starts_frame:, :]
frame_name_list = frame_name_list[starts_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list,
pos_x, pos_y, target_w, target_h,
final_score_sz, siam,
starts_frame)
lengths[idx], precisions[idx], precision_auc[idx], ious[
idx] = _compile_results(gt_, bboxes,
evaluation.dist_threshold)
print(str(i) + ' -- ' + video_list[i] + \
' -- Precision: ' + "%.2f" %precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" %precision_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
'-- Speed: ' + "%.2f" % speed[idx] + ' --\n')
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_precision_auc = np.sum(precision_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print(' -- Overall stats (averaged per frame) on ' + str(nv) +
'videos(' + str(tot_frames) + 'frames) --')
print(' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" %mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" %mean_iou +\
' -- Speed: ' + "%.2f" %mean_speed + '-- \n')
else:
gt, frame_name_list, _, _ = _init_video(env, evaluation,
evaluation.video)
pos_x, pos_y, target_w, target_h = region_to_bbox(
gt[evaluation.start_frame])
bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y,
target_w, target_h, final_score_sz, siam,
evaluation.start_frame)
_, precision, precision_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
print(evaluation.video + \
' -- precision ' + "(%d px)" %evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- precision AUC: ' + "%.2f" %precision_auc + \
' -- IOU: ' + "%.2f" %iou +\
'-- Speed: ' + "%.2f" %speed + ' -- \n')
def main():
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
siam = SiameseNet(env.root_pretrained, design.net)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder) if not v[0] == '.']
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv): # -- Iterate through all videos
print('video: %d' % (i + 1))
gt, frame_name_list, frame_sz, n_frames = _init_video(env, evaluation, videos_list[i])
starts = np.rint(np.linspace(0, n_frames - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq): # -- Iterate through a single video
start_frame = int(starts[j])
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
# bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_, pos_x, pos_y,
# target_w, target_h, final_score_sz, filename,
# image, templates_z, scores, start_frame)
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_, pos_x, pos_y,
target_w, target_h, final_score_sz, siam, start_frame) # -- here is where tracker.py is called
lengths[idx], precisions[idx], precisions_auc[idx], ious[idx] = _compile_results(gt_, bboxes, evaluation.dist_threshold)
print(str(i) + ' -- ' + videos_list[i] + \
' -- Precision: ' + "%.2f" % precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
' -- Speed: ' + "%.2f" % speed[idx] + ' --\n')
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_precision_auc = np.sum(precisions_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print('-- Overall stats (averaged per frame) on ' + str(nv) + ' videos (' + str(tot_frames) + ' frames) --')
print(' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" % mean_iou +\
' -- Speed: ' + "%.2f" % mean_speed + ' --\n')
else:
gt, frame_name_list, _, _ = _init_video(env, evaluation, evaluation.video)
pos_x, pos_y, target_w, target_h = region_to_bbox(gt[evaluation.start_frame])
# bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y, target_w, target_h, final_score_sz,
# filename, image, templates_z, scores, evaluation.start_frame)
bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y, target_w, target_h, final_score_sz,
siam, evaluation.start_frame)
_, precision, precision_auc, iou = _compile_results(gt, bboxes, evaluation.dist_threshold)
print(evaluation.video + \
' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + \
' -- Speed: ' + "%.2f" % speed + ' --\n')
def main():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
#由于得到的sorcemap与原图像大小不一致，所以要在这里按比例进行放大以此来得到图像中每个位置对应的score
#这里的final也就是design中的search_sz
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build TF graph once for all
filename, image, templates_z, scores = siam.build_tracking_graph(
final_score_sz, design, env, hp)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv):
if os.path.exists(os.path.join('data/result', videos_list[i])):
continue
gt, frame_name_list, frame_sz, n_frames, video_folder, equal = _init_video(
env, evaluation, videos_list[i])
if not equal:
print('The .jpg and .xml is not equal in', video_folder)
continue
starts = np.rint(
np.linspace(0, n_frames - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
# for j in range(evaluation.n_subseq):
for j in range(1):
start_frame = int(starts[j])
gt_ = gt[start_frame:]
frame_name_list_ = frame_name_list[start_frame:]
# pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
pos_x, pos_y, target_w, target_h = xml_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_,
pos_x, pos_y, target_w, target_h,
final_score_sz, filename, image,
templates_z, scores, start_frame)
lengths[idx], precisions[idx], precisions_auc[idx], ious[idx] =\
_compile_results(gt_, bboxes, evaluation.dist_threshold)
print str(i) + ' -- ' + videos_list[i] + \
' -- Precision: ' + "%.2f" % precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
' -- Speed: ' + "%.2f" % speed[idx] + ' --'
print
# Draw_Result(bboxes, frame_name_list_, gt_, env)
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_precision_auc = np.sum(precisions_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print
print '-- Overall stats (averaged per frame) on ' + str(
nv) + ' videos (' + str(tot_frames) + ' frames) --'
print ' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" % mean_iou +\
' -- Speed: ' + "%.2f" % mean_speed + ' --'
print
else:
gt, frame_name_list, _, _, video_folder, equal = _init_video(
env, evaluation, evaluation.video)
if not equal:
print("The .jpg and .xml is not equal in", video_folder)
exit(0)
pos_x, pos_y, target_w, target_h = xml_to_bbox(
gt[evaluation.start_frame])
bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y,
target_w, target_h, final_score_sz, filename,
image, templates_z, scores,
evaluation.start_frame)
_, precision, precision_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
print evaluation.video + \
' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + \
' -- Speed: ' + "%.2f" % speed + ' --'
print
def evaluate():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build the computational graph of Siamese fully-convolutional network
siamNet = siam.Siamese(batch_size=1)
# get tensors that will be used during tracking
image, z_crops, x_crops, templates_z, scores, loss, _, distance_to_gt, summary = siamNet.build_tracking_graph_train(
final_score_sz, design, env, hp)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv)
precisions = np.zeros(nv)
precisions_auc = np.zeros(nv)
ious = np.zeros(nv)
lengths = np.zeros(nv)
for i in range(nv):
gt, frame_name_list, frame_sz, n_frames = _init_video(
env, evaluation, videos_list[i])
gt_ = gt[0:, :]
frame_name_list_ = frame_name_list[0:]
pos_x, pos_y, target_w, target_h = region_to_bbox(
gt_[0]
) # coordinate of gt is the bottom left point of the bbox
idx = i
bboxes, speed[idx] = tracker(hp,
run,
design,
frame_name_list,
pos_x,
pos_y,
target_w,
target_h,
final_score_sz,
image,
templates_z,
scores,
path_ckpt=os.path.join(
design.saver_folder,
design.path_ckpt),
siamNet=siamNet)
lengths[idx], precisions[idx], precisions_auc[idx], ious[
idx] = _compile_results(gt_, bboxes, evaluation.dist_threshold)
print(str(i) + ' -- ' + videos_list[i] + \
' -- Precision: ' + "%.2f" % precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
' -- Speed: ' + "%.2f" % speed[idx] + ' --')
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_precision_auc = np.sum(precisions_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print('-- Overall stats (averaged per frame) on ' + str(nv) +
' videos (' + str(tot_frames) + ' frames) --')
print(' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" % mean_iou +\
' -- Speed: ' + "%.2f" % mean_speed + ' --')
#evaluate only one vedio
else:
gt, frame_name_list, frame_sz, n_frames = _init_video(
env, evaluation, evaluation.video)
pos_x, pos_y, target_w, target_h = region_to_bbox(gt[0])
bboxes, speed = tracker(hp,
run,
design,
frame_name_list,
pos_x,
pos_y,
target_w,
target_h,
final_score_sz,
image,
templates_z,
scores,
path_ckpt=os.path.join(design.saver_folder,
design.path_ckpt),
siamNet=siamNet)
_, precision, precision_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
print(evaluation.video + \
' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + \
' -- Speed: ' + "%.2f" % speed + ' --')
return precision, precision_auc, iou, speed
def main():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments(mode="conv2")
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * design.score_sz
# build TF graph once for all
image, templates_z, scores = siam.build_tracking_graph(
final_score_sz, design, env)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
# videos_list = videos_list[91:][:] #only use vot 2016
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
success_auc = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv):
# gt, frame_name_list, frame_sz, n_frames, img_mode = _init_video(env, evaluation, videos_list[i])
gt, frame_name_list, frame_sz, n_frames, img_mode = _init_video_OTB(
env, evaluation, videos_list[i])
starts = np.rint(
np.linspace(0, n_frames - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq):
start_frame = int(starts[j])
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_,
pos_x, pos_y, target_w, target_h,
final_score_sz, image,
templates_z, scores, start_frame)
lengths[idx], precisions[idx], precisions_auc[idx], ious[
idx], success_auc[idx] = _compile_results(
gt_, bboxes, evaluation.dist_threshold)
print str(i) + ' -- ' + videos_list[i] + \
' -- Precision: ' + "%.2f" % precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
' -- Success AUC: ' + "%.2f" % success_auc[idx] + \
' -- Speed: ' + "%.2f" % speed[idx] + ' --'
print
tot_frames = np.mean(lengths)
mean_precision = np.mean(precisions)
mean_precision_auc = np.mean(precisions_auc)
mean_iou = np.mean(ious)
mean_success_auc = np.mean(success_auc)
mean_speed = np.mean(speed)
print 'data set ' + evaluation.dataset + ' z_lr %f scale_step %f scale_penalty %f scale_lr %f window_influence %f' % (
hp.z_lr, hp.scale_step, hp.scale_penalty, hp.scale_lr,
hp.window_influence)
print '-- Overall stats (averaged per frame) on ' + str(
nv) + ' videos (' + str(tot_frames) + ' frames) --'
print ' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" % mean_iou +\
' -- Success AUC: ' + "%.2f" % mean_success_auc +\
' -- Speed: ' + "%.2f" % mean_speed + ' --'
print
with open('log_test.txt', 'a+') as f:
f.write(time.asctime(time.localtime(time.time())) + '\r\n')
f.write(
'data set ' + evaluation.dataset +
' z_lr %f scale_step %f scale_penalty %f scale_lr %f window_influence %f \r\n'
% (hp.z_lr, hp.scale_step, hp.scale_penalty, hp.scale_lr,
hp.window_influence))
f.write('-- Overall stats (averaged per frame) on ' + str(nv) +
' videos (' + str(tot_frames) + ' frames) --\r\n')
f.write(' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision + \
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc + \
' -- IOU: ' + "%.2f" % mean_iou + \
' -- AUC: ' + "%.3f" % mean_success_auc + \
' -- Speed: ' + "%.2f" % mean_speed + ' --\r\n')
f.write('\r\n')
else:
gt, frame_name_list, _, _ = _init_video(env, evaluation,
evaluation.video)
pos_x, pos_y, target_w, target_h = region_to_bbox(
gt[evaluation.start_frame])
bboxes, speed = tracker(hp, run, design, frame_name_list, pos_x, pos_y,
target_w, target_h, final_score_sz, filename,
image, templates_z, scores,
evaluation.start_frame)
_, precision, precision_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
print evaluation.video + \
' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + \
' -- Speed: ' + "%.2f" % speed + ' --'
print
def runSiamfc(folderPath, fourcc, testWidth, testHeight, graph, scfg):
print('Running Siamfc: ' + folderPath)
frame_name_list = _init_video(folderPath)
hp, evaluation, run, env, design = parse_arguments()
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
fp = open(os.path.join(folderPath, 'YoloBoxes.txt'))
filename, image, templates_z, scores, graph1, scfg1 = siam.build_tracking_graph(
final_score_sz, design, env)
finalImages = []
Allbboxes = []
SiamfcVid = cv2.VideoWriter(join(folderPath, 'SiamfcVid.avi'), fourcc, 10,
(testWidth, testHeight))
f = open(folderPath + "/SiamfcBoxes.txt", "w+")
nucAngles = open(os.path.join(folderPath, "nuclearAngles.txt"), "r")
print(refreshRate)
for i in range(len(frame_name_list)):
line = fp.readline()
finalImages = []
if (line == '\n'):
continue
elif (i % refreshRate == 0):
boxes = line[:-1].split(':')
boxNr = 0
for j in Allbboxes:
#label = j.label
#pos_x = j.positions[len(j.positions)-1][0]
#pos_y = j.positions[len(j.positions)-1][1]
#target_w = j.positions[len(j.positions)-1][2]
#target_h = j.positions[len(j.positions)-1][3]
#bboxes, speed, finalImages = tracker(graph1, scfg1, hp, run, design, frame_name_list[i:i+refreshRate-1], pos_x, pos_y, target_w, target_h, final_score_sz,
# filename, image, templates_z, scores, label,0,colors[boxNr%len(colors)],0,refreshRate-1, finalImages, 0)
#j.positions = np.concatenate((j.positions,bboxes),0)
j.padafter(refreshRate)
for j in boxes:
box = j.split(',')
label = box[0]
box = map(int, box[1:])
print(
'In folder %s Image %d has a box at %d,%d,%d,%d with label %s'
% (folderPath, i, box[0], box[1], box[2], box[3], label))
pos_x = box[0]
pos_y = box[1]
target_w = box[2]
target_h = box[3]
print('Pos_x: %d, Pos_y:%d, width:%d, height:%d' %
(pos_x, pos_y, target_w, target_h))
bboxes, speed, finalImages = tracker(
graph1, scfg1, hp, run, design,
frame_name_list[i:i + refreshRate - 1], pos_x, pos_y,
target_w, target_h, final_score_sz, filename, image,
templates_z, scores, label, 0, colors[boxNr % len(colors)],
0, refreshRate - 1, finalImages, 0)
newBox = sfc_bbox(colors[boxNr % len(colors)], label, bboxes,
0)
newBox.padfront(i)
Allbboxes.append(newBox)
boxNr = boxNr + 1
print(bboxes)
fname = i
probs = [0] * len(Allbboxes)
#print(Allbboxes)
if (liveFeed):
try:
oldFolderPath = folderPath.split('_')
oldFolderPath[2] = str(int(oldFolderPath[2]) - 1)
underscore = '_'
oldFolderPath = underscore.join(oldFolderPath)
except:
print("Could not load old probabilies")
oldFolderPath = ""
else:
oldFolderPath = ""
if (os.path.isfile(oldFolderPath + "/SiamfcBoxes.txt")):
print('Extracting old probs')
probs = getOldProbs(oldFolderPath, Allbboxes,
[(item.split(','))[0] for item in boxes])
print(probs)
for j in range(len(finalImages)):
#print(probs)
angle = int(nucAngles.readline())
calcProbs(finalImages[j], angle, Allbboxes, i + j, f)
cv2.circle(finalImages[j],
(int(float(1.0 - angle / 180.0) * testWidth),
int(testHeight / 2)), 10, (0, 0, 225), -1)
SiamfcVid.write(finalImages[j])
cv2.imwrite(frame_name_list[j][0:-4] + '_siamfc.png',
finalImages[j])
fname = fname + 1
#break
f.close()
SiamfcVid.release()
return
def main():
#avoid printing TF debugging information
#仅显示error log
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
#TODO:allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
#gt_,frame_name_list_,_,_=_init_video(env,evaluation,evaluation.video)
#pos_x,pos_y,target_w,target_h=region_to_bbox(gt_[0])
#print('---target_w---'+"%d"%target_w+'--target_h---'+"%d"%target_h)
#why?????????????
#Set size for use with tf.image.resize_images with align_corners=True
#For example:
# [1,4,7]=>[1 2 3 4 5 6 7] (length 3*(3-1)+1)
#instead of
#[1,4,7]=>[1 1 2 3 4 5 6 7 7](length 3*3)
#Why hp.response_up???
#design.score_sz=33
#hp.response_up=8
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
#build TF graph once for all
#filename,image,templates_z,scores are only processes.!!!
#真正返回信息需要用sess去执行（tracker中执行）
#return filename, image, templates_z, scores_up
filename, image, templates_z, scores = siam.build_tracking_graph(
final_score_sz, design, env)
#iterate through all videos of evaluation dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
#os.listdir(path):返回指定路径下的文件和文件夹
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
#遍历不同的视频样本
for i in range(nv):
#frame_name_list:each image of a video sequence
gt, frame_name_list, frame_sz, n_frames = _init_video(
env, evaluation, videos_list[i])
#np.rint():对浮点数取整但不改变浮点数类型
#n_subseq=3
starts = np.rint(
np.linspace(0, n_frame - 1, evaluation.n_subseq + 1))
#分成n_subseq+1份，将数组赋给starts
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq):
start_frame = int(starts[j])
#start_frame:指start_frame及以后(选取了n_subseq中的一份)
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
#Update
bboxes, speed[idx] = tracker(
hp,
run,
design,
env,
evaluation,
frame_name_list_,
pos_x,
pos_y,
#bboxes,speed[idx]=tracker(hp,run,design,frame_name_list_,pos_x,pos_y,
target_w,
target_h,
final_score_sz,
filename,
image,
templates_z,
scores,
start_frame)
#gt_:ground truth
#bboxes:the result of tracking
lengths[idx], precisions[idx], precisions_auc[idx], ious[
idx] = _compile_results(gt_, bboxes,
evaluation.dist_threshold)
print(
str(i) + '--' + videos_list[i] + '--Precision: ' +
"%.2f" % precisions[idx] + '--Precisions AUC: ' +
"%.2f" % precisions_auc[idx] + '--IOU: ' +
"%.2f" % ious[idx] + '--Speed: ' + "%.2f" % speed[idx] +
'--')
else:
#evaluation.video='all'
print(evaluation.video)
gt, frame_name_list, _, _ = _init_video(env, evaluation,
evaluation.video)
#evaluation.start_frame=0
pos_x, pos_y, target_w, target_h = region_to_bbox(
gt[evaluation.start_frame])
#Update
#bboxes,speed=tracker(hp,run,design,frame_name_list,pos_x,pos_y,target_w,target_h,final_score_sz,
bboxes, speed = tracker(hp, run, design, env, evaluation,
frame_name_list, pos_x, pos_y, target_w,
target_h, final_score_sz, filename, image,
templates_z, scores, evaluation.start_frame)
_, precision, precisions_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
#print(evaluation.video+
print(evaluation.video + '--Precision: ' +
"(%d px)" % evaluation.dist_threshold + ': ' +
"%.2f" % precision + '--Precisions AUC: ' +
"%.2f" % precisions_auc + '--IOU: ' + "%.2f" % iou +
'--Speed: ' + "%.2f" % speed + '--')
def main():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build TF graph once for all
filename, image, templates_z, scores = siam.build_tracking_graph(
final_score_sz, design, env)
# iterate through all videos of evaluation.dataset
if evaluation.video == 'all':
dataset_folder = os.path.join(env.root_dataset, evaluation.dataset)
videos_list = [v for v in os.listdir(dataset_folder)]
videos_list.sort()
nv = np.size(videos_list)
speed = np.zeros(nv * evaluation.n_subseq)
precisions = np.zeros(nv * evaluation.n_subseq)
precisions_auc = np.zeros(nv * evaluation.n_subseq)
ious = np.zeros(nv * evaluation.n_subseq)
lengths = np.zeros(nv * evaluation.n_subseq)
for i in range(nv):
gt, frame_name_list, frame_sz, n_frames = _init_video(
env, evaluation, videos_list[i])
starts = np.rint(
np.linspace(0, n_frames - 1, evaluation.n_subseq + 1))
starts = starts[0:evaluation.n_subseq]
for j in range(evaluation.n_subseq):
start_frame = int(starts[j])
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
idx = i * evaluation.n_subseq + j
# Call Tracker for the selected sequence
print("Tracking started!")
bboxes, speed[idx] = tracker(hp, run, design, frame_name_list_,
pos_x, pos_y, target_w, target_h,
final_score_sz, filename, image,
templates_z, scores, start_frame)
lengths[idx], precisions[idx], precisions_auc[idx], ious[
idx] = _compile_results(gt_, bboxes,
evaluation.dist_threshold)
print( str(i) + ' -- ' + videos_list[i] + \
' -- Precision: ' + "%.2f" % precisions[idx] + \
' -- Precisions AUC: ' + "%.2f" % precisions_auc[idx] + \
' -- IOU: ' + "%.2f" % ious[idx] + \
' -- Speed: ' + "%.2f" % speed[idx] + ' --' )
tot_frames = np.sum(lengths)
mean_precision = np.sum(precisions * lengths) / tot_frames
mean_precision_auc = np.sum(precisions_auc * lengths) / tot_frames
mean_iou = np.sum(ious * lengths) / tot_frames
mean_speed = np.sum(speed * lengths) / tot_frames
print('-- Overall stats (averaged per frame) on ' + str(nv) +
' videos (' + str(tot_frames) + ' frames) --')
print( ' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % mean_precision +\
' -- Precisions AUC: ' + "%.2f" % mean_precision_auc +\
' -- IOU: ' + "%.2f" % mean_iou +\
' -- Speed: ' + "%.2f" % mean_speed + ' --' )
else:
gt, frame_name_list, _, n_frames = _init_video(env, evaluation,
evaluation.video)
#pos_x, pos_y, target_w, target_h = region_to_bbox(gt[evaluation.start_frame])
# np.size(frame_name_list) = Amount of frames
# ott = amount of Objects To Track
ott = len(gt) if evaluation.multi_object else 1
objects = np.zeros((ott, 4))
for i in range(ott):
objects[i, :] = region_to_bbox(gt[i])
# Call Tracker for the selected sequence.
print("Tracking started!")
bboxes, speed = tracker(hp, run, design, frame_name_list, objects,
final_score_sz, filename, image, templates_z,
scores, evaluation.start_frame)
if evaluation.multi_object:
print('No Ground Truth available for multi object, just printing speed result....\n' + \
evaluation.video + \
' -- Speed: ' + "%.2f" % speed + ' --' )
else:
_, precision, precision_auc, iou = _compile_results(
gt, bboxes, evaluation.dist_threshold)
print( evaluation.video + \
' -- Precision ' + "(%d px)" % evaluation.dist_threshold + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + \
' -- Speed: ' + "%.2f" % speed + ' --' )
print("Tracking finished!")
def main():
# avoid printing TF debugging information
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# TODO: allow parameters from command line or leave everything in json files?
hp, evaluation, run, env, design = parse_arguments()
# Set size for use with tf.image.resize_images with align_corners=True.
# For example,
# [1 4 7] => [1 2 3 4 5 6 7] (length 3*(3-1)+1)
# instead of
# [1 4 7] => [1 1 2 3 4 5 6 7 7] (length 3*3)
final_score_sz = hp.response_up * (design.score_sz - 1) + 1
# build TF graph once for all
filename, image, templates_z, templates_x, scores, scores_original = siam.build_tracking_graph(
final_score_sz, design, env)
# create dataset
dataset = DatasetFactory.create_dataset(name=args_dataset,
dataset_root=dataset_root,
load_img=False)
# iterate through all videos of evaluation.dataset
videos_list = list(dataset.videos.keys())
videos_list.sort()
nv = np.size(videos_list)
for i in range(nv):
current_key = sorted(list(dataset.videos.keys()))[i]
gt, frame_name_list, frame_sz, n_frames = _init_video(
dataset, current_key)
for j in range(1):
start_frame = 0
gt_ = gt[start_frame:, :]
frame_name_list_ = frame_name_list[start_frame:]
pos_x, pos_y, target_w, target_h = region_to_bbox(gt_[0])
bboxes, _ = tracker(videos_list[i], hp, run, design,
frame_name_list_, pos_x, pos_y, target_w,
target_h, final_score_sz, filename, image,
templates_z, templates_x, scores,
scores_original, start_frame)
#Visualize
if visualize:
for bbox, groundt, frame_name in zip(bboxes, gt_,
frame_name_list_):
image = cv2.imread(frame_name)
bbox_pt1, bbox_pt2 = get_bbox_cv(bbox)
bbox_gt1, bbox_gt2 = get_gt_bbox_cv(groundt)
#Draw result
cv2.rectangle(image, bbox_pt1, bbox_pt2, (0, 255, 0))
#Draw ground truth
cv2.rectangle(image, bbox_gt1, bbox_gt2, (0, 0, 0))
cv2.imshow("Results:", image)
cv2.waitKey()
bboxes = bboxes.tolist()
bboxes[0] = [1]
target_dir = os.path.join(result_output, current_key)
if not os.path.exists(target_dir):
os.mkdir(target_dir)
results_file = current_key + "_" + "{:03d}".format(1) + ".txt"
results_abs_file = os.path.join(target_dir, results_file)
with open(results_abs_file, "w") as f:
for bbox in bboxes:
if len(bbox) == 1:
f.write('%d\n' % (bbox[0]))
else:
f.write('%.2f, %.2f, %.2f, %.2f\n' %
(bbox[0], bbox[1], bbox[2], bbox[3]))
