def verify_model(self):
"""
处理标记文件夹
:param img_folder: 图片文件夹
:param out_folder:
:return:
"""
yolo = YoloV3(model_path=self.model_path,
classes_path=self.classes_path,
anchors_path=self.anchors_path) # yolo算法类
img_dict = format_img_and_anno(self.img_folder)
res_dict = dict()
for count, img_name in enumerate(img_dict):
print_info('-' * 50)
print_info('图片: {}'.format(img_name))
(img_p, anno_p) = img_dict[img_name]
_, precision, recall = self.detect_img(yolo, img_p, anno_p,
self.out_folder)
res_dict[img_name] = (precision, recall)
if count == 10:
break
ap, ar = 0, 0
for name in res_dict.keys():
precision, recall = res_dict[name]
ap += precision
ar += recall
mAp = safe_div(ap, len(res_dict.keys()))
mAr = safe_div(ar, len(res_dict.keys()))
print_info('平均精准率: {:.4f} %, 平均召回率: {:.4f} %'.format(
mAp * 100, mAr * 100))
def iou_of_boxes(boxes1, boxes2):
"""
框的IOU
:param boxes1_: 真值
:param boxes2_: 预测
:return: 精准和召回
"""
res_list = []
boxes1_ = copy.deepcopy(boxes1)
boxes2_ = copy.deepcopy(boxes2)
t1, t2 = len(boxes1_), len(boxes2_)
for box1 in boxes1_:
final_iou = 0
final_box = None
for box2 in boxes2_:
iou = bb_intersection_over_union(box1, box2)
if iou > final_iou:
final_iou = iou
final_box = box2
if final_iou > 0.5:
res_list.append((box1, final_box, final_iou))
boxes2_.remove(final_box)
tr = len(res_list)
if tr == 0: # 没有检测源, 则认为正确
if t1 == 0 and t2 != 0:
recall, precision = 1.0, 0.0
elif t1 != 0 and t2 == 0:
recall, precision = 0.0, 1.0
elif t1 != 0 and t2 != 0:
recall, precision = 0.0, 0.0
else:
recall, precision = 1.0, 1.0
else:
recall = safe_div(tr, t1) # 召回率
precision = safe_div(tr, t2) # 精准率
print_info('精准: {:.4f}, 召回: {:.4f}, 匹配结果: {}'.format(
precision, recall, res_list))
return [precision, recall]
def verify_model(self):
"""
处理标记文件夹
:param img_folder: 图片文件夹
:param out_folder:
:return:
"""
img_dict = format_img_and_anno(self.img_folder)
res_list = []
for count, img_name in enumerate(img_dict):
print_info('-' * 50)
print_info('图片: {}'.format(img_name))
(img_p, anno_p) = img_dict[img_name]
try:
res_dict = self.detect_img(img_p, anno_p, self.out_folder)
except Exception as e:
print_ex('检测异常: {}'.format(e))
continue
res_list.append(res_dict)
# if count == 10:
# break
print_info('-' * 50)
for target_name in (['all'] + self.targets_name):
if target_name in ['truck', 'bus']:
continue
ap, ar, count = 0, 0, 0
for pr_dict in res_list:
if target_name in pr_dict:
tp, tr = pr_dict[target_name]
ap += tp
ar += tr
count += 1
mAp = safe_div(ap, count)
mAr = safe_div(ar, count)
print_info('类: {} P: {:.4f} %, R: {:.4f} %'.format(
target_name, mAp * 100, mAr * 100))
def iou_of_boxes(boxes1, boxes2):
res_list = []
for box1 in boxes1:
final_iou = 0
final_box = None
for box2 in boxes2:
iou = bb_intersection_over_union(box1, box2)
if iou > final_iou:
final_iou = iou
final_box = box2
if final_iou > 0.5:
res_list.append((box1, final_box, final_iou))
t1, t2, tr = len(boxes1), len(boxes2), len(res_list)
if tr == 0: # 没有检测源, 则认为正确
return [1.0, 1.0]
recall = safe_div(tr, t1) # 召回率
precision = safe_div(tr, t2) # 精准率
print_info('精准: {:.4f}, 召回: {:.4f}, 匹配结果: {}'.format(
precision, recall, res_list))
return [precision, recall]
def get_batch_acc(outputs, labels):
"""
multiLabel acc, all values are same, is right
:param outputs: predictions
:param labels: ground truth
:return: acc percent, num of right, num of all
"""
outputs = sigmoid(outputs)
outputs = outputs.asnumpy()
labels = labels.asnumpy().astype('int')
print(labels[0])
outputs = np.where(outputs > 0.5, 1, 0) # 类别阈值0.5
print(outputs[0])
rights = np.sum(np.where(labels == outputs, 0, 1), axis=1)
n_right = np.count_nonzero(rights == 0) # 全0的即全部相等
return safe_div(n_right, len(labels)), n_right, len(labels)
def main():
ym = Y3Model()
img_folder = os.path.join(IMG_DATA, 'jiaotong-0727')
right_folder = os.path.join(IMG_DATA, 'jiaotong-0727-right')
wrong_folder = os.path.join(IMG_DATA, 'jiaotong-0727-wrong')
none_folder = os.path.join(IMG_DATA, 'jiaotong-0727-none')
mkdir_if_not_exist(right_folder, is_delete=True)
mkdir_if_not_exist(wrong_folder, is_delete=True)
mkdir_if_not_exist(none_folder, is_delete=True)
img_dict = format_img_and_anno(img_folder)
r_count = 0
all_count = 0
no_recall_count = 0
for count, img_name in enumerate(img_dict):
(img_p, anno_p) = img_dict[img_name]
# print(img_p)
try:
tag_res, img_box = ym.detect_img(img_p, True)
except Exception as e:
continue
w_tags = []
for tag in tag_res.keys():
if tag_res[tag] <= 0.01:
print_info('删除Tag {} {}'.format(tag, tag_res[tag]))
w_tags.append(tag)
for tag in w_tags:
tag_res.pop(tag, None) # 小于1%的类别
all_count += 1
p_classes = set(tag_res.keys())
_, t_classes = read_anno_xml(anno_p)
merge_dict = {'truck': 'car', 'bus': 'car', 'car': 'car'} # 合并类别
t_classes = map_classes(merge_dict, t_classes) # 合并类别
traffic_names = [
'bicycle', 'car', 'motorbike', 'aeroplane', 'bus', 'train',
'truck', 'boat'
]
t_classes = set(t_classes) & set(traffic_names)
img_name = img_p.split('/')[-1]
is_right = False
if p_classes and p_classes.issubset(t_classes): # 检测正确
r_count += 1
img_box.save(os.path.join(right_folder, img_name + '.d.jpg'))
is_right = True
elif not p_classes and not t_classes: # 空，检测正确
r_count += 1
if not img_box:
img_box = Image.open(img_p)
img_box.save(os.path.join(right_folder, img_name + '.d.jpg'))
is_right = True
elif not p_classes and t_classes: # 检测为空，实际有类
if not img_box:
img_box = Image.open(img_p)
img_box.save(os.path.join(none_folder, img_name + '.d.jpg'))
no_recall_count += 1 # 未召回
r_count += 1
is_right = True
else: # 其他，检测错误
if not img_box:
img_box = Image.open(img_p)
img_box.save(os.path.join(wrong_folder, img_name + '.d.jpg'))
print_info('P: {}, T: {}, {}'.format(list(p_classes), list(t_classes),
'正确' if is_right else '错误'))
right_ratio = safe_div(r_count, all_count)
print_info('正确: {}, 全部: {}, 未召回: {}, 准确率: {}'.format(
r_count, all_count, no_recall_count, right_ratio))