def plot_gt(img, bboxes, im_path, mode='xyxyxyxy'):
if not os.path.exists('temp'):
os.mkdir('temp')
if mode == 'xywha':
bboxes = rbox_2_quad(bboxes, mode=mode)
if mode == 'xyxya':
bboxes = rbox_2_quad(bboxes, mode=mode)
for box in bboxes:
img = cv2.polylines(cv2.UMat(img),
[box.reshape(-1, 2).astype(np.int32)], True,
(0, 0, 255), 2)
cv2.imwrite(
os.path.join('temp', 'augment_%s' % (os.path.split(im_path)[1])),
img)
print('Check augmentation results in `temp` folder!!!')
def icdar_evaluate(model,
target_size,
test_path,
dataset):
if dataset == 'IC15':
output = './datasets/IC_eval/icdar15'
elif dataset == 'IC13':
output = './datasets/IC_eval/icdar13'
else:
raise NotImplementedError
ims_dir = test_path
out_dir = './temp'
if not os.path.exists(out_dir):
os.makedirs(out_dir)
ims_list = [x for x in os.listdir(ims_dir) if is_image(x)]
s = ('%20s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', 'Hmean')
nt = 0
for idx, im_name in enumerate(tqdm(ims_list, desc=s)):
im_path = os.path.join(ims_dir, im_name)
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR), cv2.COLOR_BGR2RGB)
dets = im_detect(model, im, target_sizes=target_size)
nt += len(dets)
out_file = os.path.join(out_dir, 'res_' + im_name[:im_name.rindex('.')] + '.txt')
with codecs.open(out_file, 'w', 'utf-8') as f:
if dets.shape[0] == 0:
continue
if dataset == 'IC15':
res = sort_corners(rbox_2_quad(dets[:, 2:]))
for k in range(dets.shape[0]):
f.write('{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f}\n'.format(
res[k, 0], res[k, 1], res[k, 2], res[k, 3],
res[k, 4], res[k, 5], res[k, 6], res[k, 7])
)
if dataset == 'IC13':
res = rbox_2_aabb(dets[:, 2:])
for k in range(dets.shape[0]):
f.write('{:.0f},{:.0f},{:.0f},{:.0f}\n'.format(
res[k, 0], res[k, 1], res[k, 2], res[k, 3])
)
zip_name = 'submit.zip'
make_zip(out_dir, zip_name)
shutil.move(os.path.join('./', zip_name), os.path.join(output, zip_name))
if os.path.exists(out_dir):
shutil.rmtree(out_dir)
result = os.popen('cd {0} && python script.py -g=gt.zip -s=submit.zip '.format(output)).read()
sep = result.split(':')
precision = sep[1][:sep[1].find(',')].strip()
recall = sep[2][:sep[2].find(',')].strip()
f1 = sep[3][:sep[3].find(',')].strip()
map = 0
p = eval(precision)
r = eval(recall)
hmean = eval(f1)
# display result
pf = '%20s' + '%10.3g' * 6 # print format
print(pf % ('all', len(ims_list), nt, p, r, 0, hmean))
return p, r, map, hmean
def demo(args):
#
model = STELA(backbone=args.backbone, num_classes=2)
model.load_state_dict(torch.load(args.weights))
model.eval()
ims_list = [x for x in os.listdir(args.ims_dir) if is_image(x)]
for _, im_name in enumerate(ims_list):
im_path = os.path.join(args.ims_dir, im_name)
src = cv2.imread(im_path, cv2.IMREAD_COLOR)
im = cv2.cvtColor(src, cv2.COLOR_BGR2RGB)
cls_dets = im_detect(model, im, target_sizes=args.target_size)
for j in range(len(cls_dets)):
cls, scores = cls_dets[j, 0], cls_dets[j, 1]
bbox = cls_dets[j, 2:]
if len(bbox) == 4:
draw_caption(src, bbox, '{:1.3f}'.format(scores))
cv2.rectangle(src, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])),
color=(0, 0, 255),
thickness=2)
else:
pts = np.array([rbox_2_quad(bbox[:5]).reshape((4, 2))],
dtype=np.int32)
cv2.drawContours(src, pts, 0, color=(0, 255, 0), thickness=2)
# display original anchors
# if len(bbox) > 5:
# pts = np.array([rbox_2_quad(bbox[5:]).reshape((4, 2))], dtype=np.int32)
# cv2.drawContours(src, pts, 0, color=(0, 0, 255), thickness=2)
# resize for better shown
im = cv2.resize(src, (800, 800), interpolation=cv2.INTER_LINEAR)
cv2.imshow('Detection Results', im)
cv2.waitKey(0)
def data_evaluate(model, target_size, test_path, conf=0.01, dataset=None):
root_dir = 'datasets/evaluate'
out_dir = os.path.join(root_dir, 'detection-results')
if os.path.exists(out_dir):
shutil.rmtree(out_dir)
os.makedirs(out_dir)
ds = DATASETS[dataset]()
with open(test_path, 'r') as f:
if dataset == 'VOC':
im_dir = test_path.replace('/ImageSets/Main/test.txt',
'/JPEGImages')
ims_list = [
os.path.join(im_dir,
x.strip('\n') + '.jpg') for x in f.readlines()
]
else:
ims_list = [
x.strip('\n') for x in f.readlines() if is_image(x.strip('\n'))
]
s = ('%20s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R',
'[email protected]', 'Hmean')
nt = 0
for idx, im_path in enumerate(tqdm(ims_list, desc=s)):
print(idx, im_path)
im_name = os.path.split(im_path)[1]
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
dets = im_detect(model, im, target_sizes=target_size, conf=conf)
nt += len(dets)
out_file = os.path.join(out_dir,
im_name[:im_name.rindex('.')] + '.txt')
with codecs.open(out_file, 'w', 'utf-8') as f:
if dets.shape[0] == 0:
f.close()
continue
res = sort_corners(rbox_2_quad(dets[:, 2:]))
for k in range(dets.shape[0]):
f.write(
'{} {:.2f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f}\n'
.format(ds.return_class(dets[k, 0]), dets[k, 1], res[k, 0],
res[k, 1], res[k, 2], res[k, 3], res[k, 4],
res[k, 5], res[k, 6], res[k, 7]))
# assert len(os.listdir(os.path.join(root_dir,'ground-truth'))) != 0, 'No labels found in test/ground-truth!! '
assert len(os.listdir("./HRSC2016/Test/Annotations")
) != 0, 'No labels found in test/ground-truth!! '
mAP = eval_mAP(root_dir, use_07_metric=False)
# display result
pf = '%20s' + '%10.3g' * 6 # print format
print(pf % ('all', len(ims_list), nt, 0, 0, mAP, 0))
# thres = [0.5,0.6,0.7,0.8,0.9]
# for thre in thres:
# ap = eval_mAP(root_dir, use_07_metric=True, thres = thre)
# print('ap_{}: {}'.format(thre, ap))
return 0, 0, mAP, 0
def dota_evaluate(model,
target_size,
test_path,
conf = 0.01):
#
root_data, evaldata = os.path.split(test_path)
splitdata = evaldata + 'split'
ims_dir = os.path.join(root_data, splitdata + '/' + 'images')
root_dir = 'outputs'
res_dir = os.path.join(root_dir, 'detections') # 裁剪图像的检测结果
integrated_dir = os.path.join(root_dir, 'integrated') # 将裁剪图像整合后成15个txt的结果
merged_dir = os.path.join(root_dir, 'merged') # 将整合后的结果NMS
if os.path.exists(root_dir):
shutil.rmtree(root_dir)
os.makedirs(root_dir)
for f in [res_dir, integrated_dir, merged_dir]:
if os.path.exists(f):
shutil.rmtree(f)
os.makedirs(f)
ds = DOTADataset()
# loss = torch.zeros(3)
ims_list = [x for x in os.listdir(ims_dir) if is_image(x)]
s = ('%20s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', 'Hmean')
nt = 0
for idx, im_name in enumerate(tqdm(ims_list, desc=s)):
im_path = os.path.join(ims_dir, im_name)
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR), cv2.COLOR_BGR2RGB)
dets = im_detect(model, im, target_sizes=target_size, conf = conf)
nt += len(dets)
out_file = os.path.join(res_dir, im_name[:im_name.rindex('.')] + '.txt')
with codecs.open(out_file, 'w', 'utf-8') as f:
if dets.shape[0] == 0:
f.close()
continue
res = sort_corners(rbox_2_quad(dets[:, 2:]))
for k in range(dets.shape[0]):
f.write('{:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {:.0f} {} {} {:.2f}\n'.format(
res[k, 0], res[k, 1], res[k, 2], res[k, 3],
res[k, 4], res[k, 5], res[k, 6], res[k, 7],
ds.return_class(dets[k, 0]), im_name[:-4], dets[k, 1],)
)
ResultMerge(res_dir, integrated_dir, merged_dir)
## calc mAP
mAP, classaps = task1_eval(merged_dir, test_path)
# # display result
pf = '%20s' + '%10.3g' * 6 # print format
print(pf % ('all', len(ims_list), nt, 0, 0, mAP, 0))
return 0, 0, mAP, 0
def __call__(self, img, labels, mode=None):
if random.random() < self.p:
if mode == 'xywha':
labels = rbox_2_quad(labels, mode = 'xywha')
img, labels = random_affine(img, labels,
degree=self.degree,translate=self.translate,
scale=self.scale,shear=self.shear )
labels = quad_2_rbox(labels, mode = 'xywha')
else:
img, labels = random_affine(img, labels,
degree=self.degree,translate=self.translate,
scale=self.scale,shear=self.shear )
return img, labels
def demo(args):
hyps = hyp_parse(args.hyp)
ds = DATASETS[args.dataset](level=1)
model = RetinaNet(backbone=args.backbone, hyps=hyps)
if args.weight.endswith('.pth'):
chkpt = torch.load(args.weight)
# load model
if 'model' in chkpt.keys():
model.load_state_dict(chkpt['model'])
else:
model.load_state_dict(chkpt)
print('load weight from: {}'.format(args.weight))
model.eval()
# if os.path.exists('outputs'):
# shutil.rmtree('outputs')
# os.mkdir('outputs')
t0 = time.time()
if not args.dataset == 'DOTA':
ims_list = [x for x in os.listdir(args.ims_dir) if is_image(x)]
for idx, im_name in enumerate(ims_list):
s = ''
t = time.time()
im_path = os.path.join(args.ims_dir, im_name)
s += 'image %g/%g %s: ' % (idx, len(ims_list), im_path)
src = cv2.imread(im_path, cv2.IMREAD_COLOR)
im = cv2.cvtColor(src, cv2.COLOR_BGR2RGB)
cls_dets = im_detect(model, im, target_sizes=args.target_size)
for j in range(len(cls_dets)):
cls, scores = cls_dets[j, 0], cls_dets[j, 1]
bbox = cls_dets[j, 2:]
if len(bbox) == 4:
draw_caption(src, bbox, '{:1.3f}'.format(scores))
cv2.rectangle(src, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])),
color=(0, 0, 255),
thickness=2)
else:
pts = np.array([rbox_2_quad(bbox[:5]).reshape((4, 2))],
dtype=np.int32)
cv2.drawContours(src,
pts,
0,
color=(0, 255, 0),
thickness=2)
put_label = True
if put_label:
label = ds.return_class(cls) + str(' %.2f' % scores)
fontScale = 0.7
font = cv2.FONT_HERSHEY_COMPLEX
thickness = 1
t_size = cv2.getTextSize(label,
font,
fontScale=fontScale,
thickness=thickness)[0]
c1 = tuple(bbox[:2].astype('int'))
c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 5
cv2.rectangle(src, c1, c2, [0, 255, 0], -1) # filled
cv2.putText(src,
label, (c1[0], c1[1] - 5),
font,
fontScale, [0, 0, 0],
thickness=thickness,
lineType=cv2.LINE_AA)
# display original anchors
# if len(bbox) > 5:
# pts = np.array([rbox_2_quad(bbox[5:]).reshape((4, 2))], dtype=np.int32)
# cv2.drawContours(src, pts, 0, color=(0, 0, 255), thickness=2)
# resize for better shown
# im = cv2.resize(src, (800, 800), interpolation=cv2.INTER_LINEAR)
# cv2.imshow('Detection Results', im)
# cv2.waitKey(0)
print('%sDone. (%.3fs) %d objs' %
(s, time.time() - t, len(cls_dets)))
# save image
out_path = os.path.join('outputs', os.path.split(im_path)[1])
cv2.imwrite(out_path, src)
## DOTA detct on large image
else:
evaluate(args.target_size,
args.ims_dir,
'DOTA',
args.backbone,
args.weight,
hyps=hyps,
conf=0.05)
if os.path.exists('outputs/dota_out'):
shutil.rmtree('outputs/dota_out')
os.mkdir('outputs/dota_out')
exec(
'cd outputs && rm -rf detections && rm -rf integrated && rm -rf merged'
)
ResultMerge('outputs/detections', 'outputs/integrated',
'outputs/merged', 'outputs/dota_out')
img_path = os.path.join(args.ims_dir, 'images')
label_path = 'outputs/dota_out'
save_imgs = False
if save_imgs:
show_dota_results(img_path, label_path)
print('Done. (%.3fs)' % (time.time() - t0))
def evaluate(model, args):
#
if args.dataset == 'ICDAR 2013':
ims_dir = os.path.join(args.test_dir, args.dataset, 'Task 1/Test/IMS')
eval_dir = os.path.join(args.eval_dir, 'icdar13')
elif args.dataset == 'ICDAR 2015':
ims_dir = os.path.join(args.test_dir, args.dataset, 'Task 1/Test/IMS')
eval_dir = os.path.join(args.eval_dir, 'icdar15')
elif args.dataset == 'ICDAR 2017':
ims_dir = os.path.join(args.test_dir, args.dataset, 'Task 1/Test/IMS')
eval_dir = os.path.join(args.eval_dir, 'icdar17')
elif args.dataset == 'COCO':
ims_dir = os.path.join(args.test_dir, args.dataset, 'Test/IMS')
eval_dir = os.path.join(args.eval_dir, 'coco')
else:
raise NotImplementedError
#
out_dir = './temp'
if not os.path.exists(out_dir):
os.makedirs(out_dir)
#
ims_list = [x for x in os.listdir(ims_dir) if is_image(x)]
for idx, im_name in enumerate(ims_list):
im_path = os.path.join(ims_dir, im_name)
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
_t.tic()
dets = im_detect(model, im, target_sizes=args.target_size)
print('\rim_detect: {:d}/{:d}, time: {:.3f}, '.format(
idx + 1, len(ims_list), _t.toc()),
end='')
if args.dataset == 'ICDAR 2017':
out_file = os.path.join(out_dir,
im_name[:im_name.rindex('.')] + '.txt')
out_file = out_file.replace('ts', 'res')
elif args.dataset == 'COCO':
im_index = im_name[im_name.rindex('_') + 1:im_name.rindex('.')]
out_file = os.path.join(out_dir,
'res_' + str(int(im_index)) + '.txt')
else:
out_file = os.path.join(
out_dir, 'res_' + im_name[:im_name.rindex('.')] + '.txt')
with codecs.open(out_file, 'w', 'utf-8') as f:
if dets.shape[0] == 0:
continue
if args.dataset == 'ICDAR 2013':
res = rbox_2_aabb(dets[:, 2:])
for k in range(dets.shape[0]):
f.write('{:.0f},{:.0f},{:.0f},{:.0f}\n'.format(
res[k, 0], res[k, 1], res[k, 2], res[k, 3]))
elif args.dataset == 'ICDAR 2015':
res = sort_corners(rbox_2_quad(dets[:, 2:]))
for k in range(dets.shape[0]):
f.write(
'{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f}\n'
.format(res[k, 0], res[k, 1], res[k, 2], res[k, 3],
res[k, 4], res[k, 5], res[k, 6], res[k, 7]))
elif args.dataset == 'ICDAR 2017':
res = sort_corners(rbox_2_quad(dets[:, 2:]))
for k in range(dets.shape[0]):
f.write(
'{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.0f},{:.2f}\n'
.format(res[k, 0], res[k, 1], res[k, 2], res[k, 3],
res[k, 4], res[k, 5], res[k, 6], res[k, 7],
dets[k, 1]))
if args.dataset == 'COCO':
res = rbox_2_aabb(dets[:, 2:])
for k in range(dets.shape[0]):
f.write('{:.0f},{:.0f},{:.0f},{:.0f},{:.2f}\n'.format(
res[k, 0], res[k, 1], res[k, 2], res[k, 3], dets[k,
1]))
#
zip_name = 'submit.zip'
make_zip(out_dir, zip_name)
shutil.move(os.path.join('./', zip_name), os.path.join(eval_dir, zip_name))
if os.path.exists(out_dir):
shutil.rmtree(out_dir)
if args.dataset == 'ICDAR 2013' or args.dataset == 'ICDAR 2015':
os.system(
'cd {0} && python2 script.py -g=gt.zip -s=submit.zip '.format(
eval_dir))
print()
else:
# evaluated online
raise NotImplementedError
def demo(backbone='eb2',
weights='weights/deploy_eb_ship_15.pth',
ims_dir='sample',
target_size=768):
#
model = STELA(backbone=backbone, num_classes=2)
model.load_state_dict(torch.load(weights))
# model.eval()
# print(model)
classifier = EfficientNet.from_name(net_name)
num_ftrs = classifier._fc.in_features
classifier._fc = nn.Linear(num_ftrs, class_num)
classifier = classifier.cuda()
best_model_wts = 'dataset/weismoke/model/efficientnet-b0.pth'
classifier.load_state_dict(torch.load(best_model_wts))
ims_list = [x for x in os.listdir(ims_dir) if is_image(x)]
import shutil
shutil.rmtree('output/')
os.mkdir('output/')
for _, im_name in enumerate(ims_list):
im_path = os.path.join(ims_dir, im_name)
src = cv2.imread(im_path, cv2.IMREAD_COLOR)
im = cv2.cvtColor(src, cv2.COLOR_BGR2RGB)
import time
# start=time.clock()
cls_dets = im_detect(model, im, target_sizes=target_size)
end = time.clock()
# print('********time*********',end-start)
# val='/home/jd/projects/haha/chosename/val_plane_split/label_new/'
"""
if(len(cls_dets)==0):
print('*********no********',im_name)
#image_path = os.path.join(img_path, name + ext) #样本图片的名称
shutil.move(val+im_name[0:-4]+'.txt', 'hard') #移动该样本图片到blank_img_path
shutil.move(im_path, 'hard/') #移动该样本图片的标签到blank_label_path
continue
"""
fw = open('output/' + im_name[:-4] + '.txt', 'a')
fw.truncate()
for j in range(len(cls_dets)):
cls, scores = cls_dets[j, 0], cls_dets[j, 1]
# print ('cls,score',cls,scores)
bbox = cls_dets[j, 2:]
# print(bbox)
if len(bbox) == 4:
draw_caption(src, bbox, '{:1.3f}'.format(scores))
cv2.rectangle(src, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])),
color=(0, 0, 255),
thickness=2)
else:
pts = np.array([rbox_2_quad(bbox[:5]).reshape((4, 2))],
dtype=np.int32)
# print('####pts####',pts)
cv2.drawContours(src, pts, 0, color=(0, 255, 0), thickness=2)
# display original anchors
# if len(bbox) > 5:
# pts = np.array([rbox_2_quad(bbox[5:]).reshape((4, 2))], dtype=np.int32)
# cv2.drawContours(src, pts, 0, color=(0, 0, 255), thickness=2)
patch = crop_image(im, pts)
pred = classify(classifier, patch)
fw.write(
str(pts.flatten()[0]) + ' ' + str(pts.flatten()[1]) + ' ' +
str(pts.flatten()[2]) + ' ' + str(pts.flatten()[3]) + ' ' +
str(pts.flatten()[4]) + ' ' + str(pts.flatten()[5]) + ' ' +
str(pts.flatten()[6]) + ' ' + str(pts.flatten()[7]) + ' ' +
classes[pred] + '\n')
fw.close()
# resize for better shown
im = cv2.resize(src, (768, 768), interpolation=cv2.INTER_LINEAR)
cv2.imwrite('output_img/' + im_name, im)
train_img_dir = '/home/jd/projects/bifpn/sample_plane/'
groundtruth_txt_dir = '/home/jd/projects/haha/chosename/val_plane_split/label_new/'
detect_txt_dir = '/home/jd/projects/bifpn/output/'
Recall, Precision, mAP = compute_acc(train_img_dir, groundtruth_txt_dir,
detect_txt_dir)
print('*******', Recall)
return Recall, Precision, mAP
