def track_video(model, video):
toc, regions = 0, []
image_files, gt = video['image_files'], video['gt']
for f, image_file in enumerate(image_files):
im = cv2.imread(image_file) # TODO: batch load
tic = cv2.getTickCount()
if f == 0: # init
target_pos, target_sz = rect_2_cxy_wh(gt[f])
state = SiamRPN_init(im, target_pos, target_sz,
model) # init tracker
location = cxy_wh_2_rect(state['target_pos'], state['target_sz'])
regions.append(gt[f])
elif f > 0: # tracking
state = SiamRPN_track(state, im) # track
location = cxy_wh_2_rect(state['target_pos'] + 1,
state['target_sz'])
regions.append(location)
toc += cv2.getTickCount() - tic
if args.visualization and f >= 0: # visualization
if f == 0:
cv2.destroyAllWindows()
if len(gt[f]) == 8:
cv2.polylines(im,
[np.array(gt[f], np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
else:
cv2.rectangle(im, (gt[f, 0], gt[f, 1]),
(gt[f, 0] + gt[f, 2], gt[f, 1] + gt[f, 3]),
(0, 255, 0), 3)
if len(location) == 8:
cv2.polylines(im, [location.reshape((-1, 1, 2))], True,
(0, 255, 255), 3)
else:
location = [int(l) for l in location] #
cv2.rectangle(
im, (location[0], location[1]),
(location[0] + location[2], location[1] + location[3]),
(0, 255, 255), 3)
cv2.putText(im, str(f), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1,
(0, 255, 255), 2)
cv2.imshow(video['name'], im)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save result
video_path = join('../test', args.dataset, 'SiamRPN_AlexNet_OTB2015')
if not isdir(video_path): makedirs(video_path)
result_path = join(video_path, '{:s}.txt'.format(video['name']))
with open(result_path, "w") as fin:
for x in regions:
fin.write(','.join([str(i) for i in x]) + '\n')
print('({:d}) Video: {:12s} Time: {:02.1f}s Speed: {:3.1f}fps'.format(
v_id, video['name'], toc, f / toc))
return f / toc
def main(imagedir, gtdir):
# load net
net_file = join(realpath(dirname(__file__)), 'SiamRPNBIG.model')
net = SiamRPNBIG()
net.load_state_dict(torch.load(net_file))
net.eval().cuda()
# warm up
for i in range(10):
net.temple(
torch.autograd.Variable(torch.FloatTensor(1, 3, 127, 127)).cuda())
net(torch.autograd.Variable(torch.FloatTensor(1, 3, 255, 255)).cuda())
# start to track
# get the first frame groundtruth
gt_file = os.path.join(gtdir, 'gt.txt')
with open(gt_file, 'r') as f:
lines = f.readlines()
gt = []
for line in lines:
line = line.split(' ')
gt.append([int(float(x)) for x in line])
init_bbox = gt[0] # top-left x y,w,h
target_pos, target_sz = rect_2_cxy_wh(
init_bbox) # top-left x y,w,h --> center x y,w,h
image_list = glob.glob(os.path.join(imagedir, '*.jpg'))
image_list.sort()
im = cv2.imread(image_list[0]) # HxWxC
state = SiamRPN_init(im, target_pos, target_sz, net) # init tracker
bboxes = []
for i in range(1, len(gt)):
im = cv2.imread(image_list[i]) # HxWxC
state = SiamRPN_track(state, im) # track
res = cxy_wh_2_rect(
state['target_pos'],
state['target_sz']) # center x y,w,h --> top-left x y,w,h
bboxes.append(res.tolist())
_, precision, precision_auc, iou = _compile_results(gt[1:], bboxes)
print(' -- Precision ' + "(20 px)" + ': ' + "%.2f" % precision +\
' -- Precision AUC: ' + "%.2f" % precision_auc + \
' -- IOU: ' + "%.2f" % iou + ' --')
isSavebbox = True
if isSavebbox:
print('saving bbox...')
res_bbox_file = os.path.join('results_bbox.json')
json.dump(bboxes, open(res_bbox_file, 'w'), indent=2)
isSavevideo = True
if isSavevideo:
print('saving video...')
save_video(image_list, bboxes)
print('done')
for jj in range(int(_data_provider.cur_img_num / 100)):
exemplar_list = [None for i in range(batch_size)]
source_list = [None for i in range(batch_size)]
instance_list = [None for i in range(batch_size)]
exemplar_cxy_list = [[None, None] for i in range(batch_size)]
source_cxy_list = [[None, None] for i in range(batch_size)]
instance_cxy_list = [[None, None] for i in range(batch_size)]
for batch in range(batch_size):
pairs, gts = _data_provider.rand_pick_pair()
exemplar = cv2.imread(pairs[0])
source = cv2.imread(pairs[1])
instance = cv2.imread(pairs[2])
exemplar_pos, exemplar_sz = rect_2_cxy_wh(gts[0])
source_pos, source_sz = rect_2_cxy_wh(gts[1])
instance_pos, instance_sz = rect_2_cxy_wh(gts[2])
exemplar_list[batch] = exemplar
source_list[batch] = source
instance_list[batch] = instance
exemplar_cxy_list[batch][0], exemplar_cxy_list[batch][
1] = exemplar_pos, exemplar_sz
source_cxy_list[batch][0], source_cxy_list[batch][
1] = source_pos, source_sz
instance_cxy_list[batch][0], instance_cxy_list[batch][
1] = instance_pos, instance_sz
train_config = SiamRPN_init_batch(exemplar_list, exemplar_cxy_list,
net)
total_failure = 0
warped_images = []
video_length = vot.get_frame_length(video_name)
#ground truth bounding box
gts = vot.get_gts(video_name)
frame_tags = vot.get_frame_tags(video_name)
video_frames = vot.get_frames(video_name)
flow_dir = os.path.join(flow_dirs, video_name + '.txt')
img_dir = os.path.join(vot_dir, video_name,'color')
confidence_dir = os.path.join('/home/jianingq/backward_flow_confidence_vot/',video_name)
#initialize network
# image and init box
init_rbox = gts[0]
if(len(init_rbox) == 4):
[cx, cy], [w, h] = rect_2_cxy_wh(init_rbox)
else:
[cx, cy, w, h] = get_axis_aligned_bbox(init_rbox)
# tracker init
target_pos, target_sz = np.array([cx, cy]), np.array([w, h])
im = video_frames[0]# HxWxC
state = SiamRPN_init(im, target_pos, target_sz, net)
detection_box = [int(cx-w/2),int(cy-h/2),int(cx+w/2),int(cy+h/2)]
for i in range(0,video_length - 1):
#track
im1 = np.copy(video_frames[i])
im2 = np.copy(video_frames[i + 1])
entropy_data = np.load(os.path.join(confidence_dir,format(i+1, '08')+'_entropy.npy'))
def rtaa_attack(net,
x_init,
x,
gt,
target_pos,
target_sz,
scale_z,
p,
eps=10,
alpha=1,
iteration=10,
x_val_min=0,
x_val_max=255):
x = Variable(x.data)
x_adv = Variable(x_init.data, requires_grad=True)
alpha = eps * 1.0 / iteration
for i in range(iteration):
delta, score = net(x_adv)
score_temp = score.permute(1, 2, 3, 0).contiguous().view(2, -1)
score = torch.transpose(score_temp, 0, 1)
delta1 = delta.permute(1, 2, 3, 0).contiguous().view(4, -1)
delta = delta.permute(1, 2, 3,
0).contiguous().view(4, -1).data.cpu().numpy()
# calculate proposals
gt_cen = rect_2_cxy_wh(gt)
gt_cen = np.tile(gt_cen, (p.anchor.shape[0], 1))
gt_cen[:, 0] = ((gt_cen[:, 0] - target_pos[0]) * scale_z -
p.anchor[:, 0]) / p.anchor[:, 2]
gt_cen[:, 1] = ((gt_cen[:, 1] - target_pos[1]) * scale_z -
p.anchor[:, 1]) / p.anchor[:, 3]
gt_cen[:, 2] = np.log(gt_cen[:, 2] * scale_z) / p.anchor[:, 2]
gt_cen[:, 3] = np.log(gt_cen[:, 3] * scale_z) / p.anchor[:, 3]
# create pseudo proposals randomly
gt_cen_pseudo = rect_2_cxy_wh(gt)
gt_cen_pseudo = np.tile(gt_cen_pseudo, (p.anchor.shape[0], 1))
rate_xy1 = np.random.uniform(0.3, 0.5)
rate_xy2 = np.random.uniform(0.3, 0.5)
rate_wd = np.random.uniform(0.7, 0.9)
gt_cen_pseudo[:, 0] = ((gt_cen_pseudo[:, 0] - target_pos[0] -
rate_xy1 * gt_cen_pseudo[:, 2]) * scale_z -
p.anchor[:, 0]) / p.anchor[:, 2]
gt_cen_pseudo[:, 1] = ((gt_cen_pseudo[:, 1] - target_pos[1] -
rate_xy2 * gt_cen_pseudo[:, 3]) * scale_z -
p.anchor[:, 1]) / p.anchor[:, 3]
gt_cen_pseudo[:, 2] = np.log(
gt_cen_pseudo[:, 2] * rate_wd * scale_z) / p.anchor[:, 2]
gt_cen_pseudo[:, 3] = np.log(
gt_cen_pseudo[:, 3] * rate_wd * scale_z) / p.anchor[:, 3]
delta[0, :] = (delta[0, :] * p.anchor[:, 2] +
p.anchor[:, 0]) / scale_z + target_pos[0]
delta[1, :] = (delta[1, :] * p.anchor[:, 3] +
p.anchor[:, 1]) / scale_z + target_pos[1]
delta[2, :] = (np.exp(delta[2, :]) * p.anchor[:, 2]) / scale_z
delta[3, :] = (np.exp(delta[3, :]) * p.anchor[:, 3]) / scale_z
location = np.array([
delta[0] - delta[2] / 2, delta[1] - delta[3] / 2, delta[2],
delta[3]
])
label = overlap_ratio(location, gt)
# set thresold to define positive and negative samples, following the training step
iou_hi = 0.6
iou_low = 0.3
# make labels
y_pos = np.where(label > iou_hi, 1, 0)
y_pos = torch.from_numpy(y_pos).cuda().long()
y_neg = np.where(label < iou_low, 0, 1)
y_neg = torch.from_numpy(y_neg).cuda().long()
pos_index = np.where(y_pos.cpu() == 1)
neg_index = np.where(y_neg.cpu() == 0)
index = np.concatenate((pos_index, neg_index), axis=1)
# make pseudo lables
y_pos_pseudo = np.where(label > iou_hi, 0, 1)
y_pos_pseudo = torch.from_numpy(y_pos_pseudo).cuda().long()
y_neg_pseudo = np.where(label < iou_low, 1, 0)
y_neg_pseudo = torch.from_numpy(y_neg_pseudo).cuda().long()
y_truth = y_pos
y_pseudo = y_pos_pseudo
# calculate classification loss
loss_truth_cls = -F.cross_entropy(score[index], y_truth[index])
loss_pseudo_cls = -F.cross_entropy(score[index], y_pseudo[index])
loss_cls = (loss_truth_cls - loss_pseudo_cls) * (1)
# calculate regression loss
loss_truth_reg = -rpn_smoothL1(delta1, gt_cen, y_pos)
loss_pseudo_reg = -rpn_smoothL1(delta1, gt_cen_pseudo, y_pos)
loss_reg = (loss_truth_reg - loss_pseudo_reg) * (5)
# final adversarial loss
loss = loss_cls + loss_reg
# calculate the derivative
net.zero_grad()
if x_adv.grad is not None:
x_adv.grad.data.fill_(0)
loss.backward(retain_graph=True)
adv_grad = where((x_adv.grad > 0) | (x_adv.grad < 0), x_adv.grad, 0)
adv_grad = torch.sign(adv_grad)
x_adv = x_adv - alpha * adv_grad
x_adv = where(x_adv > x + eps, x + eps, x_adv)
x_adv = where(x_adv < x - eps, x - eps, x_adv)
x_adv = torch.clamp(x_adv, x_val_min, x_val_max)
x_adv = Variable(x_adv.data, requires_grad=True)
return x_adv
net.load_state_dict(model_dict)
net.cuda().eval()
# warm up
for i in range(10):
net.temple(torch.autograd.Variable(torch.ones(1, 3, 127, 127)).cuda(), \
torch.autograd.Variable(torch.ones(1, 3, 271, 271)).cuda())
net(torch.autograd.Variable(torch.ones(1, 3, 271, 271)).cuda())
for seq in OTB_seqs:
_data_provider.pick_seq(seq)
exemplar_path, exemplar_gt, cur_img_num = _data_provider.eval_pick_exemplar(
)
exemplar = cv2.imread(exemplar_path)
exemplar_pos, exemplar_sz = rect_2_cxy_wh(exemplar_gt)
state = SiamRPN_init(exemplar, exemplar_pos, exemplar_sz, net)
save_file = save_res_path + seq + '_ours.txt'
tracking_res = open(save_file, 'w')
for idx in range(cur_img_num):
instance_path = _data_provider.eval_pick_instance()
instance = cv2.imread(instance_path)
state = SiamRPN_track(state, instance)
print('seq:{}:{} , score:{}'.format(seq, idx, state['score']))
res = cxy_wh_2_rect(state['target_pos'], state['target_sz'])
tracking_res.write('{},{},{},{}'.format(res[0], res[1], res[2],
res[3]))
tracking_res.write('\n')