def init(self, im, init_rbox):
state = self.state
[cx, cy, w, h] = get_axis_aligned_bbox(init_rbox)
# tracker init
target_pos, target_sz = np.array([cx, cy]), np.array([w, h])
p = TrackerConfig()
p.update(self.net.cfg)
state['im_h'] = im.shape[0]
state['im_w'] = im.shape[1]
if p.adaptive:
if ((target_sz[0] * target_sz[1]) /
float(state['im_h'] * state['im_w'])) < 0.004:
p.instance_size = 287 # small object big search region
else:
p.instance_size = 271
#python3
p.score_size = (p.instance_size -
p.exemplar_size) // p.total_stride + 1 #python3
p.anchor = generate_anchor(p.total_stride, p.scales, p.ratios,
p.score_size)
avg_chans = np.mean(im, axis=(0, 1))
wc_z = target_sz[0] + p.context_amount * sum(target_sz)
hc_z = target_sz[1] + p.context_amount * sum(target_sz)
s_z = Round(np.sqrt(wc_z * hc_z)) #python3
# initialize the exemplar
z_crop = get_subwindow_tracking(im, target_pos, p.exemplar_size, s_z,
avg_chans)
z_crop = Variable(z_crop.unsqueeze(0))
if self.step == 1: #不更新模板
self.net.temple(z_crop.cuda()) #初始化模板
else: #更新模板
z_f = self.net.featextract(z_crop.cuda()) #[1,512,6,6]
self.net.kernel(z_f)
# state['z_0'] = z_f.cpu().data #初始的模板
state['z_f_cur'] = z_f.cpu().data #当前检测特征图
state['z_f_pre'] = z_f.cpu().data #前一帧的检测特征图
# state['z_f'] = z_f.cpu().data #累积的模板
# state['z_a'] = z_f.cpu().data #累积的模板
state['gt_f_cur'] = z_f.cpu().data #gt框对应的特征图
# z_f = z_f.repeat(self.cfg["UPDATE"]["NUM_FRAMES"],1,1,1)
# z_f = self.feat_to_input(z_f)
# state['z_f_frames'] = z_f.cpu().data
state['z_f'] = z_f.cpu().data #累积的模板
state['z_0'] = z_f.cpu().data
# state['success'] = 1
# 111 112 123
# 取之前三帧template作为输入
# pre_idx = [index - pre_idx_diff]
# for idx in range(self.cfg["UPDATE"].get("NUM_FRAMES", 3)-1):
# # self.train_data["pre"][pre_idx[idx]] could be 0
# # in this case, all frames before that would be the initial frame
# offset = self.train_data["pre"][pre_idx[idx]] if idx < init_idx_diff - 1 else 0
# pre_idx += [pre_idx[idx] - offset]
# pre_idx.sort()
# template_frames = torch.tensor(self.train_data["template_cur"][pre_idx])
# # T, C, H, W -> C, T, H, W
# template_frames = template_frames.permute(1,0,2,3)
if p.windowing == 'cosine':
window = np.outer(np.hanning(p.score_size),
np.hanning(p.score_size))
elif p.windowing == 'uniform':
window = np.ones((p.score_size, p.score_size))
window = np.tile(window.flatten(), p.anchor_num)
state['p'] = p
state['net'] = self.net
state['avg_chans'] = avg_chans
state['window'] = window
state['target_pos'] = target_pos
state['target_sz'] = target_sz
self.state = state
return state
def validation(self, im, state=None, given_cls=1):
if state == None:
state = self.state
p = state['p']
net = state['net']
avg_chans = state['avg_chans']
window = state['window']
target_pos = state['target_pos']
target_sz = state['target_sz']
if 'gt_pos' in state.keys():
gt_pos = state['gt_pos']
gt_sz = state['gt_sz']
wc_z = gt_sz[1] + p.context_amount * sum(gt_sz)
hc_z = gt_sz[0] + p.context_amount * sum(gt_sz)
s_z = np.sqrt(wc_z * hc_z) #2020-05-13
gt_crop = Variable(
get_subwindow_tracking(im, gt_pos, p.exemplar_size, Round(s_z),
avg_chans).unsqueeze(0))
g_f = net.featextract(gt_crop.cuda())
state['gt_f_cur'] = g_f.cpu().data #当前帧gt框对应的特征模板
wc_z = target_sz[1] + p.context_amount * sum(target_sz)
hc_z = target_sz[0] + p.context_amount * sum(target_sz)
s_z = np.sqrt(wc_z * hc_z)
scale_z = p.exemplar_size / s_z
d_search = (p.instance_size - p.exemplar_size) / 2 #python3 2020-05-13
pad = d_search / scale_z
s_x = s_z + 2 * pad
# extract scaled crops for search region x at previous target position
#x_crop = Variable(get_subwindow_tracking(im, target_pos, p.instance_size, round(s_x), avg_chans).unsqueeze(0))
x_c = get_subwindow_tracking(im, target_pos, p.instance_size,
Round(s_x), avg_chans)
x_crop = Variable(x_c.unsqueeze(0))
resmap_pure = self.net.resmap_fc(x_crop.cuda(), state['z_f'].cuda())
target_pos, target_sz, score = self.tracker_eval(
x_crop.cuda(), target_pos, target_sz * scale_z, window, scale_z, p)
target_pos[0] = max(0, min(state['im_w'], target_pos[0]))
target_pos[1] = max(0, min(state['im_h'], target_pos[1]))
target_sz[0] = max(10, min(state['im_w'], target_sz[0]))
target_sz[1] = max(10, min(state['im_h'], target_sz[1]))
#更新模板
if self.step > 1:
z_crop = Variable(
get_subwindow_tracking(im, target_pos, p.exemplar_size,
Round(s_z), avg_chans).unsqueeze(0))
z_f = self.net.featextract(z_crop.cuda()) #当前检测模板
state['z_f_cur'] = z_f.cpu().data
if self.step == 2: #模板更新方式1-Linear
zLR = 0.0102 #SiamFC[0.01, 0.05], 0.0102是siamfc初始化的方法
z_f_ = (1 - zLR) * Variable(
state['z_f']).cuda() + zLR * z_f # 累积模板
else: #模板更新方式2-UpdateNet
init_inp = state['z_0'].cuda()
acc_inp = state['z_f'].cuda()
z_f_, xcls = self.updatenet.validation(
[z_f, init_inp, acc_inp], given_cls)
if xcls[0, 0] > xcls[0, 1]:
success = 1
else:
success = 0
# z_f_, success = self.updatenet.predict([z_f, init_inp, acc_inp, state['z_f_pre'].cuda()])
state['z_f'] = z_f_.cpu().data #累积模板
state['z_f_pre'] = state['z_f_cur']
self.net.kernel(z_f_) #更新模板
state['success'] = success
state['target_pos'] = target_pos
state['target_sz'] = target_sz
state['score'] = score
# state['z_f'] = z_f_.cpu().data #累积模板
self.state = state
return state
def update(self, im):
state = self.state
p = state['p']
net = state['net']
avg_chans = state['avg_chans']
window = state['window']
target_pos = state['target_pos']
target_sz = state['target_sz']
wc_z = target_sz[1] + p.context_amount * sum(target_sz)
hc_z = target_sz[0] + p.context_amount * sum(target_sz)
s_z = np.sqrt(wc_z * hc_z)
scale_z = p.exemplar_size / s_z
d_search = (p.instance_size - p.exemplar_size) / 2 #python3 2020-05-13
pad = d_search / scale_z
s_x = s_z + 2 * pad
# extract scaled crops for search region x at previous target position
#x_crop = Variable(get_subwindow_tracking(im, target_pos, p.instance_size, round(s_x), avg_chans).unsqueeze(0))
x_c = get_subwindow_tracking(im, target_pos, p.instance_size,
Round(s_x), avg_chans)
x_crop = Variable(x_c.unsqueeze(0))
target_pos, target_sz, score = self.tracker_eval(
x_crop.cuda(), target_pos, target_sz * scale_z, window, scale_z, p)
target_pos[0] = max(0, min(state['im_w'], target_pos[0]))
target_pos[1] = max(0, min(state['im_h'], target_pos[1]))
target_sz[0] = max(10, min(state['im_w'], target_sz[0]))
target_sz[1] = max(10, min(state['im_h'], target_sz[1]))
#更新模板
if self.step > 1:
z_crop = Variable(
get_subwindow_tracking(im, target_pos, p.exemplar_size,
Round(s_z), avg_chans).unsqueeze(0))
z_f = self.net.featextract(z_crop.cuda()) #检测模板
if self.step == 2: #模板更新方式1-Linear
zLR = 0.0102 #SiamFC[0.01, 0.05], 0.0102是siamfc初始化的方法
z_f_ = (1 - zLR) * Variable(
state['z_f']).cuda() + zLR * z_f # 累积模板
else: #模板更新方式2-UpdateNet
temp = torch.cat((Variable(
state['z_0']).cuda(), Variable(state['z_f']).cuda(), z_f),
1)
#temp = np.concatenate((init, pre, cur), axis=1)
if self.step == 4:
init_inp = Variable(state['z_0']).cuda()
z_f_ = self.updatenet(temp, init_inp)
else:
z_f_ = self.updatenet(temp)
state['z_f'] = z_f_.cpu().data #累积模板
self.net.kernel(z_f_) #更新模板
state['target_pos'] = target_pos
state['target_sz'] = target_sz
state['score'] = score
self.state = state
return state