def coarse_location(self, hp_score_up, p_score_up, scale_score, ltrbs):
upsize = (cfg.TRACK.SCORE_SIZE - 1) * cfg.TRACK.STRIDE + 1
max_r_up_hp, max_c_up_hp = np.unravel_index(hp_score_up.argmax(), hp_score_up.shape) # [rpw, col]
max_r = int(round(max_r_up_hp / scale_score))
max_c = int(round(max_c_up_hp / scale_score))
# max_r = bbox_clip(max_r, 0, cfg.TRACK.SCORE_SIZE)# [0,25]超出范围则截断处理
# max_c = bbox_clip(max_c, 0, cfg.TRACK.SCORE_SIZE)# [0,25]超出范围则截断处理
max_r = bbox_clip(max_r, 0, cfg.TRACK.SCORE_SIZE-1)# 下标从0开始，这里最大值取不到SCORE_SIZE所以应该SCORE_SIZE-1，原始开源代码并没有-1操作，也没有报错，为何？
max_c = bbox_clip(max_c, 0, cfg.TRACK.SCORE_SIZE-1)# 下标从0开始，这里最大值取不到SCORE_SIZE所以应该SCORE_SIZE-1
bbox_region = ltrbs[max_r, max_c, :] #找到最大值对应的回归分支的坐标
min_bbox = int(cfg.TRACK.REGION_S * cfg.TRACK.EXEMPLAR_SIZE) # 127*0.1=12
max_bbox = int(cfg.TRACK.REGION_L * cfg.TRACK.EXEMPLAR_SIZE) # 127*0.44=55
l_region = int(min(max_c_up_hp, bbox_clip(bbox_region[0], min_bbox, max_bbox)) / 2.0) #？？
t_region = int(min(max_r_up_hp, bbox_clip(bbox_region[1], min_bbox, max_bbox)) / 2.0) #？？
r_region = int(min(upsize - max_c_up_hp, bbox_clip(bbox_region[2], min_bbox, max_bbox)) / 2.0)
b_region = int(min(upsize - max_r_up_hp, bbox_clip(bbox_region[3], min_bbox, max_bbox)) / 2.0)
# 注意这里的掩码操作
mask = np.zeros_like(p_score_up)
mask[max_r_up_hp - t_region:max_r_up_hp + b_region + 1, max_c_up_hp - l_region:max_c_up_hp + r_region + 1] = 1
p_score_up = p_score_up * mask
return p_score_up
def track(self, img):
"""
args:
img(np.ndarray): BGR image
return:
bbox(list):[x, y, width, height]
"""
w_z = self.size[0] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
h_z = self.size[1] + cfg.TRACK.CONTEXT_AMOUNT * np.sum(self.size)
s_z = np.sqrt(w_z * h_z)
self.scale_z = cfg.TRACK.EXEMPLAR_SIZE / s_z
s_x = s_z * (cfg.TRACK.INSTANCE_SIZE / cfg.TRACK.EXEMPLAR_SIZE)
x_crop = self.get_subwindow(img,
self.center_pos, cfg.TRACK.INSTANCE_SIZE,
round(s_x), self.channel_average)
# 【1， 3， 255， 255】
outputs = self.model.track(
x_crop
) # [cls, loc, cen] --> [[1, 2, 25,25], [1, 4, 25,25], [1,1,25,25]]
cls = self._convert_cls(
outputs['cls']).squeeze() # [25, 25] .data.cpu().numpy()
cen = outputs['cen'].data.cpu().numpy() # [1,1,25,25]
cen = (cen -
cen.min()) / cen.ptp() #x.ptp()对所有数据求最大值和最小值的差值，归一化到[0,1]之间
cen = cen.squeeze() #[25, 25]
ltrbs = outputs['loc'].data.cpu().numpy().squeeze(
) #[1, 4, 25, 25] --> [4, 25, 25]
upsize = (cfg.TRACK.SCORE_SIZE -
1) * cfg.TRACK.STRIDE + 1 #（25-1）× 8 +1 = 193 ？？？
penalty = self.cal_penalty(ltrbs, cfg.TRACK.PENALTY_K)
p_score = penalty * cls * cen
if cfg.TRACK.hanming:
hp_score = p_score * (1 - cfg.TRACK.WINDOW_INFLUENCE
) + self.window * cfg.TRACK.WINDOW_INFLUENCE
else:
hp_score = p_score
#25 25
hp_score_up = cv2.resize(
hp_score, (upsize, upsize),
interpolation=cv2.INTER_CUBIC) # 线性插值 [193, 193]
p_score_up = cv2.resize(
p_score, (upsize, upsize),
interpolation=cv2.INTER_CUBIC) # 线性插值 [193, 193]
cls_up = cv2.resize(cls, (upsize, upsize),
interpolation=cv2.INTER_CUBIC) # 线性插值 [193, 193]
ltrbs = np.transpose(ltrbs, (1, 2, 0)) # [4, 25, 25] --> [25,25,4]
ltrbs_up = cv2.resize(
ltrbs, (upsize, upsize),
interpolation=cv2.INTER_CUBIC) #线性插值 [25, 25, 4] --> [193, 193,4]
scale_score = upsize / cfg.TRACK.SCORE_SIZE # SCORE_SIZE=25
# get center
max_r_up, max_c_up, new_cx, new_cy = self.getCenter(
hp_score_up, p_score_up, scale_score, ltrbs)
# get w h
ave_w = (ltrbs_up[max_r_up, max_c_up, 0] +
ltrbs_up[max_r_up, max_c_up, 2]) / self.scale_z
ave_h = (ltrbs_up[max_r_up, max_c_up, 1] +
ltrbs_up[max_r_up, max_c_up, 3]) / self.scale_z
# 计算惩罚权重
s_c = self.change(
self.sz(ave_w, ave_h) /
self.sz(self.size[0] * self.scale_z, self.size[1] * self.scale_z))
r_c = self.change((self.size[0] / self.size[1]) / (ave_w / ave_h))
penalty = np.exp(-(r_c * s_c - 1) * cfg.TRACK.PENALTY_K)
# 惩罚因子×分类得分×学习率=新的学习率
lr = penalty * cls_up[max_r_up, max_c_up] * cfg.TRACK.LR #
new_width = lr * ave_w + (
1 - lr) * self.size[0] #对宽和高进行不同程度的调节 这里宽和高的学习率是否可以改成动态的？？
new_height = lr * ave_h + (1 - lr) * self.size[1]
# clip boundary
cx = bbox_clip(new_cx, 0, img.shape[1])
cy = bbox_clip(new_cy, 0, img.shape[0])
width = bbox_clip(new_width, 0, img.shape[1])
height = bbox_clip(new_height, 0, img.shape[0])
# udpate state
self.center_pos = np.array([cx, cy])
self.size = np.array([width, height])
bbox = [cx - width / 2, cy - height / 2, width, height]
return {
'bbox': bbox,
}