def read_label(self, label_file, names_dict, anchors):
'''
parsement label_file, and generates label_y1, label_y2, label_y3
return:label_y1, label_y2, label_y3
'''
contents = tools.parse_voc_xml(label_file, names_dict)
if not contents:
return None, None, None
# flip the label
if self.is_flip:
self.is_flip = False
for i in range(len(contents)):
contents[i][1] = 1.0 - contents[i][1]
label_y1 = np.zeros((self.height // 32, self.width // 32, 3, 5 + self.class_num), np.float32)
label_y2 = np.zeros((self.height // 16, self.width // 16, 3, 5 + self.class_num), np.float32)
label_y3 = np.zeros((self.height // 8, self.width // 8, 3, 5 + self.class_num), np.float32)
delta = self.smooth_delta
if delta:
label_y1[:, :, :, 4] = delta / self.class_num
label_y2[:, :, :, 4] = delta / self.class_num
label_y3[:, :, :, 4] = delta / self.class_num
y_true = [label_y3, label_y2, label_y1]
ratio = {0:8, 1:16, 2:32}
test_result = []
for label in contents:
label_id = int(label[0])
box = np.asarray(label[1: 5]).astype(np.float32) # the value saved in label is x,y,w,h
test_result.append([box[0]-box[2]/2, box[1]-box[3]/2, box[0]+box[2]/2, box[1]+box[3]/2])
best_giou = 0
best_index = 0
for i in range(len(anchors)):
min_wh = np.minimum(box[2:4], anchors[i])
max_wh = np.maximum(box[2:4], anchors[i])
giou = (min_wh[0] * min_wh[1]) / (max_wh[0] * max_wh[1])
if giou > best_giou:
best_giou = giou
best_index = i
# 012->0, 345->1, 678->2
x = int(np.floor(box[0] * self.width / ratio[best_index // 3]))
y = int(np.floor(box[1] * self.height / ratio[best_index // 3]))
k = best_index % 3
y_true[best_index // 3][y, x, k, 0:4] = box
# label smooth
label_value = 1.0 if not delta else (1-delta)
y_true[best_index // 3][y, x, k, 5:] = delta/self.class_num
y_true[best_index // 3][y, x, k, 4:5] = label_value
y_true[best_index // 3][y, x, k, 5 + label_id] = label_value
return label_y1, label_y2, label_y3, test_result
def parse_voc_xml(self, xml_file):
'''
return: ids and boxes are integrate list
'''
# 这里的box的 [name_id, xmin, ymin, xmax, ymax] 是整数
contents = tools.parse_voc_xml(xml_file,
names_dict=self.names_dict,
get_ori=True)
ids = []
boxes = []
for content in contents:
ids.append(content[0])
boxes.append(content[1:])
return ids, boxes
def read_label(self, label_file, names_dict):
'''
读取 label_file, 并生成 label_y1, label_y2, label_y3
return:label_y1, label_y2, label_y3
'''
contents = tools.parse_voc_xml(label_file, names_dict)
if not contents:
return None, None, None
label_y1 = np.zeros((self.height // 32, self.width // 32, 3, 5 + self.class_num), np.float32)
label_y2 = np.zeros((self.height // 16, self.width // 16, 3, 5 + self.class_num), np.float32)
label_y3 = np.zeros((self.height // 8, self.width // 8, 3, 5 + self.class_num), np.float32)
y_true = [label_y3, label_y2, label_y1]
ratio = {0:8, 1:16, 2:32}
for label in contents:
label_id = int(label[0])
box = np.asarray(label[1: 5]).astype(np.float32) # label中保存的就是 x,y,w,h
best_giou = 0
best_index = 0
for i in range(len(self.anchors)):
min_wh = np.minimum(box[2:4], self.anchors[i])
max_wh = np.maximum(box[2:4], self.anchors[i])
giou = (min_wh[0] * min_wh[1]) / (max_wh[0] * max_wh[1])
if giou > best_giou:
best_giou = giou
best_index = i
# 012->0, 345->1, 678->2
x = int(np.floor(box[0] * self.width / ratio[best_index // 3]))
y = int(np.floor(box[1] * self.height / ratio[best_index // 3]))
k = best_index % 3
y_true[best_index // 3][y, x, k, 0:4] = box
y_true[best_index // 3][y, x, k, 4:5] = 1.0
y_true[best_index // 3][y, x, k, 5 + label_id] = 1.0
return label_y1, label_y2, label_y3
def read_label(self, label_file, names_dict, anchors, new_w, new_h):
'''
读取 label_file, 并生成 label_y1, label_y2, label_y3
new_w:缩放以后的图片宽,真值
new_h:缩放以后的图片高,真值
return:label_y1, label_y2, label_y3
'''
contents = tools.parse_voc_xml(label_file, names_dict)
if not contents:
return None, None, None
# flip the label
if self.flip_img:
for i in range(len(contents)):
contents[i][1] = 1.0 - contents[i][1]
if self.keep_img_shape:
x_pad = (self.width - new_w) // 2
y_pad = (self.height - new_h) // 2
label_y1 = np.zeros(
(self.height // 32, self.width // 32, 3, 5 + self.class_num),
np.float32)
label_y2 = np.zeros(
(self.height // 16, self.width // 16, 3, 5 + self.class_num),
np.float32)
label_y3 = np.zeros(
(self.height // 8, self.width // 8, 3, 5 + self.class_num),
np.float32)
y_true = [label_y3, label_y2, label_y1]
ratio = {0: 8, 1: 16, 2: 32}
test_result = []
for label in contents:
label_id = int(label[0])
box = np.asarray(label[1:5]).astype(
np.float32) # label中保存的就是 x,y,w,h
if self.keep_img_shape:
# 加入填充的黑边宽高，重新修正坐标
box[0:2] = (box[0:2] * [new_w, new_h] +
[x_pad, y_pad]) / [self.width, self.height]
box[2:4] = (box[2:4] *
[new_w, new_h]) / [self.width, self.height]
test_result.append([
box[0] - box[2] / 2, box[1] - box[3] / 2, box[0] + box[2] / 2,
box[1] + box[3] / 2
])
best_giou = 0
best_index = 0
for i in range(len(anchors)):
min_wh = np.minimum(box[2:4], anchors[i])
max_wh = np.maximum(box[2:4], anchors[i])
giou = (min_wh[0] * min_wh[1]) / (max_wh[0] * max_wh[1])
if giou > best_giou:
best_giou = giou
best_index = i
# 012->0, 345->1, 678->2
x = int(np.floor(box[0] * self.width / ratio[best_index // 3]))
y = int(np.floor(box[1] * self.height / ratio[best_index // 3]))
k = best_index % 3
y_true[best_index // 3][y, x, k, 0:4] = box
# label smooth
label_value = 1.0 if not self.label_smooth else (
(1 - self.smooth_delta) +
self.smooth_delta * 1 / self.class_num)
y_true[best_index // 3][y, x, k, 4:5] = label_value
y_true[best_index // 3][
y, x, k, 5:
-1] = 0.0 if not self.label_smooth else self.smooth_delta / self.class_num
y_true[best_index // 3][y, x, k, 5 + label_id] = label_value
return label_y1, label_y2, label_y3, test_result