def eval_rotatexml(GT_xml_dir, det_xml_dir, NAME_LABEL_MAP,
file_ext='.xml', ovthresh=0.5):
# GT_xml_path 是需要转换为txt的xml文件
# txt_dir_h是将要写入的xml转换为txt的文件路径
# annopath 是GT的txt文件
# 读取原图路径
LABEl_NAME_MAP = get_label_name_map(NAME_LABEL_MAP)
file_paths = get_file_paths_recursive(GT_xml_dir, '.xml')
angle_det_all=[]
for count, xml_path in enumerate(file_paths):
img_size,gsd,imagesource,gtbox_label,extra=read_rotate_xml(xml_path,NAME_LABEL_MAP)
det_xml=xml_path.replace(GT_xml_dir,det_xml_dir)
try:
img_size,gsd,imagesource,detbox_label,extra=read_rotate_xml(det_xml,NAME_LABEL_MAP)
except :
continue
cvrboxes=[]
for box in gtbox_label:
cvrboxes.append(rotate_rect2cv_np(box))
for box in detbox_label:
cvrboxes.append(rotate_rect2cv_np(box))
cvrboxes = np.array(cvrboxes)
score=np.ones((len(gtbox_label)))
score_det=np.array(extra)
score=np.hstack((score,score_det))
# gtbox_label=np.array(gtbox_label+detbox_label)
if len(cvrboxes)>0:##斜框NMS
keep ,angle_det= nms_rotate_cpu(cvrboxes, score, ovthresh,
200) #这里可以改
angle_det=np.array(angle_det)
inx=angle_det>0
angle_det=angle_det[inx]
angle_det=angle_det.tolist()
# assert(len(angle_det)==len(gtbox_label))
print(len(angle_det))
angle_det_all+=angle_det
print(len(angle_det_all))
angle_det_all=np.array(angle_det_all)
mean_angle=np.mean(angle_det_all)
print('Mean of angle is {}'.format(mean_angle))
GT_xml_path = '/media/zf/U/2021ZKXT_aug/annotations/valrotatexml'
txt_dir_h = '/media/zf/U/2021ZKXT_aug/annotations/GT'
imagesetfile = osp.join(osp.dirname(GT_xml_path), 'gt_list.txt')
generate_file_list(GT_xml_path, imagesetfile, file_ext='.xml')
# LABEl_NAME_MAP = get_label_name_map(NAME_LABEL_MAP)
LABEl_NAME_MAP_eval = get_label_name_map(NAME_LABEL_MAP_eval)
file_paths = get_file_paths_recursive(GT_xml_path, '.xml')
if os.path.isdir(txt_dir_h):
shutil.rmtree(txt_dir_h, True)
os.makedirs(txt_dir_h)
### GT
for count, xml_path in enumerate(file_paths):
img_size, gsd, imagesource, gtbox_label, extra = read_rotate_xml(
xml_path, NAME_LABEL_MAP)
# eval txt
CLASS_DOTA = NAME_LABEL_MAP.keys()
# Task1 #
write_handle_h = open(
os.path.join(
txt_dir_h,
'{}.txt'.format(os.path.splitext(os.path.split(xml_path)[1])[0])),
'w') #Task1_gt_
# gtbox_label=np.array(gtbox_label)
ploys = rotated_box_to_poly_np(gtbox_label)
for i, rect_box in enumerate(ploys):
# rbox[4]=0
# rbox_cv=rotate_rect2cv(rbox)
def merge_rotatexml(xmldir, merge_xmldir, NAME_LABEL_MAP, NAME_LONG_MAP,
std_scale, score_threshold):
LABEL_NAME_MAP = get_label_name_map(NAME_LABEL_MAP)
if not os.path.exists(merge_xmldir):
os.mkdir(merge_xmldir)
imgs_path = get_file_paths_recursive(xmldir, '.xml')
name_det = {}
for num, img_path in enumerate(imgs_path, 0):
str_num = img_path.replace('.xml', '').strip().split('__')
img_name = str_num[0]
scale = np.float(str_num[1])
ww_ = np.int(str_num[2])
hh_ = np.int(str_num[3])
img_size, gsd, imagesource, gtbox, extra = read_rotate_xml(
img_path, NAME_LABEL_MAP)
gtbox = np.array(gtbox)
if len(gtbox) > 0:
gtbox[:, 6] = extra
inx = gtbox[:, 6] > score_threshold
rotateboxes = gtbox[inx]
gtbox[:, 0] = (gtbox[:, 0] + ww_) / scale
gtbox[:, 1] = (gtbox[:, 1] + hh_) / scale
gtbox[:, 2] /= scale
gtbox[:, 3] /= scale
if not img_name in name_det:
name_det[img_name] = gtbox
else:
name_det[img_name] = np.vstack((name_det[img_name], gtbox))
# else:
# print(' ')
# # #将所有检测结果综合起来 #正框mns
for img_name, rotateboxes in name_det.items():
if std_scale == 0:
print('no nms')
else:
inx, scores = py_cpu_label_nms(
dets=np.array(rotateboxes, np.float32),
thresh=nms_threshold,
max_output_size=500,
LABEL_NAME_MAP=LABEL_NAME_MAP_USnavy_20,
NAME_LONG_MAP=NAME_LONG_MAP_USnavy,
std_scale=std_scale)
rotateboxes = rotateboxes[inx]
# result[:, 4]=scores[inx]
# inx=result[:, 4]>score_threshold_2
# result=result[inx]
cvrboxes = []
for box in rotateboxes:
cvrboxes.append(rotate_rect2cv_np(box))
cvrboxes = np.array(cvrboxes)
if len(cvrboxes) > 0: ##斜框NMS
keep = nms_rotate_cpu(cvrboxes, rotateboxes[:, 6], rnms_threshold,
4000) #这里可以改
rotateboxes = rotateboxes[keep]
# keep=[]#去掉过 尺寸异常的目标
# for i in range(rotateboxes.shape[0]):
# box=rotateboxes[i,:]
# actual_long=box[3]*scale
# standad_long = NAME_LONG_MAP[LABEL_NAME_MAP[box[5]]]
# STD_long = NAME_LONG_MAP[LABEL_NAME_MAP[box[5]]]
# STD_long *= std_scale
# if np.abs(actual_long-standad_long)/standad_long < STD_long *1.2 and box[2]>16:
# keep.append(i)
# else:
# print('{} hh:{} ww:{}'.format(img_name,hh_,ww_))
# print('{} th label {} is wrong,long is {} normal long is {} width is {}'.format(i+1,LABEl_NAME_MAP[box[5]],actual_long,standad_long,box[3]))
# rotateboxes=rotateboxes[keep]
#保存检测结果 为比赛格式
# image_dir,image_name=os.path.split(img_path)
# gt_box=rotateboxes[:,0:5]
# label=rotateboxes[:,6]
write_rotate_xml(merge_xmldir, '{}.jpg'.format(img_name),
[1024, 1024, 3], 0.5, 'USnavy', rotateboxes,
LABEL_NAME_MAP, rotateboxes[:,
6]) #size,gsd,imagesource
def eval_rotatexml(GT_xml_dir, det_xml_dir, NAME_LABEL_MAP,
file_ext='.xml', ovthresh=0.5):
# GT_xml_path 是需要转换为txt的xml文件
# txt_dir_h是将要写入的xml转换为txt的文件路径
# annopath 是GT的txt文件
# 读取原图路径
LABEl_NAME_MAP = get_label_name_map(NAME_LABEL_MAP)
file_paths = get_file_paths_recursive(GT_xml_dir, '.xml')
num_det_all=[]
num_tp_all=[]
num_gt_all=[]
for count, xml_path in enumerate(file_paths):
img_size,gsd,imagesource,gtbox_label,extra=read_rotate_xml(xml_path,NAME_LABEL_MAP)
# num_gt = np.ones((len(gtbox_label)))
num_det = np.zeros((len(gtbox_label)))
num_tp = np.zeros((len(gtbox_label)))
det_xml=xml_path.replace(GT_xml_dir,det_xml_dir)
try:
img_size,gsd,imagesource,detbox_label,extra=read_rotate_xml(det_xml,NAME_LABEL_MAP)
except :
continue
if len(detbox_label)>0:
for i,box in enumerate( gtbox_label):
gt_cvrbox=rotate_rect2cv_np(box)
gt_center=np.array([box[0],box[1]])
gt_centers=np.repeat(gt_center[None], len(detbox_label), axis=0)
det_centers=np.array(detbox_label)[:,0:2]
diff=det_centers-gt_centers
dist=np.sqrt(np.square(diff[:,0])+np.square(diff[:,1]),)
index=np.argmin(dist)
det_gt_box=detbox_label[index]
det_cvrbox=rotate_rect2cv_np(det_gt_box)
r1 = ((gt_cvrbox[0], gt_cvrbox[1]), (gt_cvrbox[2], gt_cvrbox[3]), gt_cvrbox[4])
area_r1 = gt_cvrbox[2] * gt_cvrbox[3]#计算当前检测框的面积
r2 = ((det_cvrbox[0], det_cvrbox[1]), (det_cvrbox[ 2], det_cvrbox[3]), det_cvrbox[4])
area_r2 = det_cvrbox[2] * det_cvrbox[3]
inter = 0.0
int_pts = cv2.rotatedRectangleIntersection(r1, r2)[1]#求两个旋转矩形的交集，并返回相交的点集合
if int_pts is not None:
order_pts = cv2.convexHull(int_pts, returnPoints=True)#求点集的凸边形
int_area = cv2.contourArea(order_pts)#计算当前点集合组成的凸边形的面积
inter = int_area * 1.0 / (area_r1 + area_r2 - int_area + 0.0000001)
if inter >= ovthresh:#对大于设定阈值的检测框进行滤除
num_det[i]=1
# angle_diff=np.min([np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]),
# np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]+np.pi*2),
# np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]-np.pi*2),
# np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]+np.pi),
# np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]-np.pi)])
angle_diff=np.min([np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]),
np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]+np.pi*2),
np.abs(gtbox_label[i][4]-detbox_label[int(index)][4]-np.pi*2)])
if angle_diff<np.pi/18:
num_tp[i]=1
else:
print('xmlpath {}'.format(xml_path))
print('gt index {}'.format(i))
print('det index {}'.format(index))
num_det=num_det.tolist()
num_tp=num_tp.tolist()
# num_gt=num_gt.tolist()
# assert(len(angle_det)==len(gtbox_label))
print(len(num_det))
num_det_all+=num_det
num_tp_all+=num_tp
print(len(num_det_all))
num_det_all=np.array(num_det_all)
mean_det=np.mean(num_det_all)
num_tp_all=np.array(num_tp_all)
mean_tp=np.mean(num_tp_all)
print('Mean of det is {}'.format(mean_det))
print('Mean of gt is {}'.format(mean_tp))
print('bow acc rate {}'.format(mean_tp/mean_det))