def siamese_init(im, target_pos, target_sz, model, hp=None, device='cpu'):
"""
初始化跟踪器，根据目标的信息构建state 字典
:param im: 当前处理的图像
:param target_pos: 目标的位置
:param target_sz: 目标的尺寸
:param model: 训练好的网络模型
:param hp: 超参数
:param device: 硬件信息
:return: 跟踪器的state字典数据
"""
# 初始化state字典
state = dict()
# 设置图像的宽高
state['im_h'] = im.shape[0]
state['im_w'] = im.shape[1]
# 配置跟踪器的相关参数
p = TrackerConfig()
# 对参数进行更新
p.update(hp, model.anchors)
# 更新参数
p.renew()
# 获取网络模型
net = model
# 根据网络参数对跟踪器的参数进行更新，主要是anchors
p.scales = model.anchors['scales']
p.ratios = model.anchors['ratios']
p.anchor_num = model.anchor_num
# 生成锚点
p.anchor = generate_anchor(model.anchors, p.score_size)
# 图像的平均值
avg_chans = np.mean(im, axis=(0, 1))
# 根据设置的上下文比例，输入z 的宽高及尺寸
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))
# 初始化跟踪目标 initialize the exemplar
z_crop = get_subwindow_tracking(im, target_pos, p.exemplar_size, s_z,
avg_chans)
# 将其转换为Variable可在pythorch中进行反向传播
z = Variable(z_crop.unsqueeze(0))
# 专门处理模板
net.template(z.to(device))
# 设置使用的惩罚窗口
if p.windowing == 'cosine':
# 利用hanning窗的外积生成cosine窗口
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))
# 每一个anchor都有一个对应的惩罚窗口
window = np.tile(window.flatten(), p.anchor_num)
# 将信息更新到state字典中
state['p'] = p
state['net'] = net
state['avg_chans'] = avg_chans
state['window'] = window
state['target_pos'] = target_pos
state['target_sz'] = target_sz
return state
def siamese_init(im, target_pos, target_sz, model, hp=None, device='cpu'):
# print("------siamese_init-------")
state = dict()
state['im_h'] = im.shape[0]
state['im_w'] = im.shape[1]
# print("im.shape[0] ", im.shape[0])
p = TrackerConfig()
p.update(hp, model.anchors)
p.renew()
net = model
p.scales = model.anchors['scales']
p.ratios = model.anchors['ratios']
p.anchor_num = model.anchor_num
p.anchor = generate_anchor(model.anchors, 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))
# initialize the exemplar
z_crop = get_subwindow_tracking(im, target_pos, p.exemplar_size, s_z,
avg_chans)
# print("z size (patch) ", z_crop.size())
z = Variable(z_crop.unsqueeze(0))
# La xarxa es guarda les features resultants (self.zf) d'haver passat el patch z per la siamesa
net.template(z.to(device))
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'] = net
state['avg_chans'] = avg_chans
state['window'] = window
state['target_pos'] = target_pos
state['target_sz'] = target_sz
# print("window = ", state['window'])
return state
def siamese_init(im, target_pos, target_sz, model, hp=None, device='cpu'):
state = dict()
state['im_h'] = im.shape[0]
state['im_w'] = im.shape[1]
p = TrackerConfig()
p.update(hp, model.anchors)
p.renew()
net = model
p.scales = model.anchors['scales']
p.ratios = model.anchors['ratios']
p.anchor_num = model.anchor_num
p.anchor = generate_anchor(model.anchors, 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))
# initialize the exemplar
z_crop = get_subwindow_tracking(im, target_pos, p.exemplar_size, s_z,
avg_chans)
z = Variable(z_crop.unsqueeze(0))
net.template(z.to(device))
if p.windowing == 'cosine':
window = np.outer(np.hanning(p.score_size),
np.hanning(p.score_size)) # 外积的结果是矩阵，内积的结果是一个数
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'] = net
state['avg_chans'] = avg_chans
state['window'] = window
state['target_pos'] = target_pos
state['target_sz'] = target_sz
return state
def siamese_init(im,
model,
hp=None,
device='cpu',
targets=None,
detector=None):
custom_objects = detector.CustomObjects(car=True, person=True)
state = dict()
state['im_h'] = im.shape[0]
state['im_w'] = im.shape[1]
p = TrackerConfig()
p.update(hp, model.anchors)
p.renew()
net = model
p.scales = model.anchors['scales']
p.ratios = model.anchors['ratios']
p.anchor_num = model.anchor_num
p.anchor = generate_anchor(model.anchors, p.score_size)
avg_chans = np.mean(im, axis=(0, 1))
# s_z = [ round(np.sqrt(target["target_sz"][1] + 0.123 * sum(target["target_sz"])*target["target_sz"][0] + 0.123 * sum(target["target_sz"]) )) for target in targets ]
# s_z = np.array(s_z)
# print(targe)
# targets.append(targe)
# print(targets)
BLUE = [255, 255, 255]
for i, target in enumerate(targets):
wc_z = target["target_sz"][0] + p.context_amount * sum(
target["target_sz"])
hc_z = target["target_sz"][1] + p.context_amount * sum(
target["target_sz"])
target["s_z"] = round(np.sqrt(wc_z * hc_z))
print("out")
# initialize the exemplar
targets = get_subwindow_tracking(
im,
p.exemplar_size,
avg_chans,
targets=targets,
)
# z_f = [ net.template(Variable(target["im_to_torch"].unsqueeze(0)).to(device)) for target in targets ]
for i, target in enumerate(targets):
# detections = detector.detectCustomObjectsFromImage(custom_objects=custom_objects, input_image=target["im_patch"],input_type="array", output_image_path=os.path.join("image {} custom.jpg".format(i)),output_type="file", minimum_percentage_probability=30)
# detections = detector.detectCustomObjectsFromImage(custom_objects=custom_objects, input_image=target["img"],input_type="array", output_image_path=os.path.join(execution_path , "images.jpg"),output_type="file", minimum_percentage_probability=30)
z = Variable(target["im_to_torch"].unsqueeze(0))
target["zf"] = net.template(z.to(device))
del target["im_to_torch"]
# for eachObject in detections:
# print(eachObject["name"] , " : ", eachObject["percentage_probability"], " : ", eachObject["box_points"] )
# target["detection"] = eachObject["box_points"]
# print("--------------------------------")
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'] = net
state['avg_chans'] = avg_chans
state['window'] = window
state["targets"] = targets
state["detector"] = detector
# state["s_z"] = s_z
# state["z_f"] = z_f
return state
def siamese_init(im,
search_shape,
target_pos,
target_sz,
model,
hp=None,
device='cpu'):
"""
generate anchors, inference the template image, set up window
:param im: whole image
:param target_pos: target position that are selected
:param target_sz: target size that are selected
:param model: SiamMask model
:param hp: hyper parameters
:param device:
:return:
"""
state = dict()
state['im_h'] = search_shape[0]
state['im_w'] = search_shape[1]
p = TrackerConfig()
p.update(hp, model.anchors)
p.renew()
net = model
p.scales = model.anchors['scales']
p.ratios = model.anchors['ratios']
p.anchor_num = model.anchor_num
p.anchor = generate_anchor(
model.anchors, p.score_size) # anchor size: (25*25*5, 4) --> (3125, 4)
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)) # crop size = sqrt((w+(w+h)/2)*(h+(w+h)/2))
## initialize the exemplar
#im_patch = get_subwindow_tracking(im, target_pos, p.exemplar_size, s_z, avg_chans, out_mode="numpy")
im_patch = im
im_patch = cv2.resize(im_patch, (p.exemplar_size, p.exemplar_size))
cv2.imshow('crop_template', im_patch)
cv2.waitKey(0)
z_crop = im_to_torch(im_patch)
z = Variable(z_crop.unsqueeze(0))
net.template(z.to(device))
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'] = net
state['avg_chans'] = avg_chans
state['window'] = window
state['target_pos'] = target_pos
state['target_sz'] = target_sz
return state
def siamese_init(im,
target_pos,
target_sz,
model,
hp=None,
device='cpu'): #target_pos, target_sz输入的就是由gt轴对称得来的
state = dict()
state['im_h'] = im.shape[0]
state['im_w'] = im.shape[1]
p = TrackerConfig() #配置参数
p.update(hp,
model.anchors) #用hp,model.anchors更新p的参数，相当于用config_vot.json更新p
p.renew()
# p.score_size=25
net = model
p.scales = model.anchors['scales'] #Custom的父类SiamMask里的属性
p.ratios = model.anchors['ratios']
p.anchor_num = model.anchor_num #vot数据集上是5
p.anchor = generate_anchor(
model.anchors, p.score_size
) #generate_anchor 生成锚点。p.anchor.shape = (p.anchor_num*p.score_size*p.score_size,4)
avg_chans = np.mean(im, axis=(0, 1)) #此处im单张图片，对每个颜色通道都求均值（3，）（B,G,R)
#图像预处理，按比例外扩目标框，从而获得一定的 context 信息。p.context_amount = 0.5
wc_z = target_sz[0] + p.context_amount * sum(
target_sz) #wc_z = w + p.context_amuont * (w+h)
hc_z = target_sz[1] + p.context_amount * sum(
target_sz) #hc_z = h + p.context_amuont * (w+h)
#需要将框定的框做一个大约2倍放大,以物体为中心， s_z为宽高，截取一个正方体的物体出来
s_z = round(np.sqrt(wc_z *
hc_z)) #round四舍五入取整，round(2.5) = 2,round(2.51)=3
# initialize the exemplar
z_crop = get_subwindow_tracking(
im, target_pos, p.exemplar_size, s_z,
avg_chans) #tensor<(3, 127, 127), float32, cpu>
#TrackerConfig中的定义是 input z size，127
#z_crop的维度是(127*127*3)
z = Variable(
z_crop.unsqueeze(0)
) #pytorch中的命令，扩充数据维度,变成神经网络的参数tensor<(1, 3, 127, 127), float32, cpu>
net.template(z.to(device)) #将z送到cuda上面提取特征，即得到resnet50之后的结果
if p.windowing == 'cosine': #默认
window = np.outer(
np.hanning(p.score_size), np.hanning(p.score_size)
) #求外积 ndarray(p.score_size，p.score_size)即<(25, 25), float64>
elif p.windowing == 'uniform':
window = np.ones((p.score_size, p.score_size))
window = np.tile(window.flatten(),
p.anchor_num) #对window.flatten()在X轴进行重复p.anchor_num次
#ndarray<(3125,), float64>,p.anchor_num=5
state['p'] = p
state['net'] = net
state['avg_chans'] = avg_chans
state['window'] = window
state['target_pos'] = target_pos #还是传进来的数据
state['target_sz'] = target_sz #还是传进来的数据
return state