def detect(net, img_path, thresh, wid, hei):
#img = cv2.imread(img_path, cv2.IMREAD_COLOR)
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
if wid > 0 and hei > 0:
image = cv2.resize(img, (wid, hei))
else:
max_im_shrink = np.sqrt(1700 * 1200 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
with torch.no_grad():
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy()
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
conf = "{:.3f}".format(score)
point = (int(left_up[0]), int(left_up[1] - 5))
# cv2.putText(img, conf, point, cv2.FONT_HERSHEY_COMPLEX,
# 0.6, (0, 255, 0), 1)
t2 = time.time()
#print('detect:{} timer:{}'.format(img_path, t2 - t1))
cv2.imwrite(os.path.join(args.save_dir, os.path.basename(img_path)), img)
return t2 - t1
def detect_face(net, img, thresh):
height, width, _ = img.shape
x = to_chw_bgr(img)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
bboxes = []
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
box = []
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(np.int)
j += 1
box += [pt[0], pt[1], pt[2] - pt[0], pt[3] - pt[1], score]
bboxes += [box]
return bboxes
def detect_face(net, img, shrink):
if shrink != 1:
img = cv2.resize(img,
None,
None,
fx=shrink,
fy=shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(img)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
# print(x.size())
y = net(x)
detections = y.data
detections = detections.cpu().numpy()
det_conf = detections[0, 1, :, 0]
det_xmin = img.shape[1] * detections[0, 1, :, 1] / shrink
det_ymin = img.shape[0] * detections[0, 1, :, 2] / shrink
det_xmax = img.shape[1] * detections[0, 1, :, 3] / shrink
det_ymax = img.shape[0] * detections[0, 1, :, 4] / shrink
det = np.column_stack((det_xmin, det_ymin, det_xmax, det_ymax, det_conf))
keep_index = np.where(det[:, 4] >= args.thresh)[0]
det = det[keep_index, :]
return det
def detect_face(net, img, thresh):
height, width, _ = img.shape
im_shrink = 640.0 / max(height, width)
image = cv2.resize(img, None, None, fx=im_shrink,
fy=im_shrink, interpolation=cv2.INTER_LINEAR).copy()
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
y = net(x)
detections = y.data
scale = torch.Tensor([img.shape[1], img.shape[0],
img.shape[1], img.shape[0]])
bboxes = []
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
box = []
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(np.int)
j += 1
box += [pt[0], pt[1], pt[2], pt[3], score]
box[1] += 0.2 * (box[3] - box[1] + 1)
bboxes += [box]
return bboxes
def detect(net, img_path, thresh):
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(2000 * 2000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
# x = x * cfg.scale
# scaling
if cfg.rescale:
x = 1 / 255. * (x + 127.5)
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
detect_results = []
for i in range(detections.size(1)):
j = 0
while j < detections.size(2) and detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
detect_results.append([pt[0], pt[1], pt[2], pt[3], float(score)])
j += 1
detect_results = np.array(detect_results)
if len(detect_results.shape) == 1:
return np.array([])
order = detect_results[:, 4].ravel().argsort()[::-1]
det = detect_results[order, :]
t2 = time.time()
print('detect:{} timer:{}'.format(img_path, t2 - t1))
return det
def detect(net, img_path, thresh):
img = Image.open(img_path)
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
if os.environ['IS_QVGA_MONO'] == '1':
max_im_shrink = np.sqrt(320 * 240 / (img.shape[0] * img.shape[1]))
else:
max_im_shrink = np.sqrt(640 * 480 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
# img = cv2.resize(img, (640, 640))
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
if cfg.IS_MONOCHROME == True:
x = 0.299 * x[0] + 0.587 * x[1] + 0.114 * x[2]
x = torch.from_numpy(x).unsqueeze(0).unsqueeze(0)
else:
x = torch.from_numpy(x).unsqueeze(0)
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy()
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
conf = "{:.3f}".format(score)
point = (int(left_up[0]), int(left_up[1] - 5))
cv2.putText(img, conf, point, cv2.FONT_HERSHEY_COMPLEX, 0.6,
(0, 255, 0), 1)
t2 = time.time()
print('detect:{} timer:{}'.format(img_path, t2 - t1))
cv2.imwrite(os.path.join(args.save_dir, os.path.basename(img_path)), img)
def detect(net, img_path, thresh):
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1500 * 1000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
cnt_faces = 0
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
cnt_faces += 1
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
conf = "{:.2f}".format(score)
text_size, baseline = cv2.getTextSize(conf,
cv2.FONT_HERSHEY_SIMPLEX,
0.3, 1)
p1 = (left_up[0], left_up[1] - text_size[1])
cv2.rectangle(img, (p1[0] - 2 // 2, p1[1] - 2 - baseline),
(p1[0] + text_size[0], p1[1] + text_size[1]),
[255, 0, 0], -1)
cv2.putText(img, conf, (p1[0], p1[1] + baseline),
cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 255, 255), 1, 8)
t2 = time.time()
print('detect:{}, num_faces:{}, timer:{}'.format(img_path, cnt_faces,
t2 - t1))
cv2.imwrite(os.path.join(args.save_dir, os.path.basename(img_path)), img)
def detect(net, img_path, thresh):
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(
1500 * 1000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img, None, None, fx=max_im_shrink,
fy=max_im_shrink, interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor([img.shape[1], img.shape[0],
img.shape[1], img.shape[0]])
# 图片类型
img_type = img_path.split('.')[-1]
# 原图
origin_img = Image.open(img_path)
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
# score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
# left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
xmin, ymin, xmax, ymax = pt[0], pt[1], pt[2], pt[3]
xmin = max(0, xmin-5)
ymin = max(0, ymin-5)
xmax = min(width, xmax+5)
ymax = min(height, ymax+5)
crop_img = origin_img.crop((xmin, ymin, xmax, ymax))
crop_img = crop_img.convert('RGB')
global name_count
name_count += 1
filename = (name_ss + str(name_count))[-6:]
print(args.save_dir + filename + img_type)
crop_img.save(args.save_dir + filename + '.' + img_type)
j += 1
def detect(net, img_path, thresh, save_crops):
img = Image.open(img_path)
#orig_img = img
#if img.mode == 'L':
# img = img.convert('RGB')
#orig_img = np.array(orig_img)
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1200 * 1100 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
#x = x * cfg.scale
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
for i in range(detections.size(1)):
j = 0
image_filename = os.path.join(args.save_dir,
os.path.basename(img_path))
while detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
if save_crops:
# increase bbox by 10% on all sides
face_width = pt[3] - pt[1]
ten_percent_width = int(0.1 * face_width)
face_height = pt[2] - pt[0]
ten_percent_height = int(0.1 * face_height)
# save cropped face images as face0-imagename, face1-imagename, etc.
face = img[pt[1] - ten_percent_width:pt[3] + ten_percent_width,
pt[0] - ten_percent_height:pt[2] +
ten_percent_height]
cv2.imwrite(
os.path.join(
args.save_dir,
'face{}-'.format(j) + os.path.basename(img_path)),
cv2.cvtColor(face, cv2.COLOR_RGB2BGR))
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
conf = "{:.2f}".format(score)
text_size, baseline = cv2.getTextSize(conf,
cv2.FONT_HERSHEY_SIMPLEX,
0.3, 1)
p1 = (left_up[0], left_up[1] - text_size[1])
cv2.rectangle(img, (p1[0] - 2 // 2, p1[1] - 2 - baseline),
(p1[0] + text_size[0], p1[1] + text_size[1]),
[255, 0, 0], -1)
cv2.putText(img, conf, (p1[0], p1[1] + baseline),
cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 255, 255), 1, 8)
t2 = time.time()
print('detect:{} timer:{}'.format(img_path, t2 - t1))
cv2.imwrite(image_filename, cv2.cvtColor(img, cv2.COLOR_RGB2BGR))
def detect(net, im, thresh, pre_bbox):
#img = Image.open(img_path)
img = Image.fromarray(cv2.cvtColor(im, cv2.COLOR_BGR2RGB))
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1200 * 1000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
#image = cv2.resize(img,(224,),interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
#t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
#img = cv2.imread(img_path, cv2.IMREAD_COLOR)
mask_img = np.ones(im.shape, np.int8)
kernel_size = 15
blur_img = cv2.blur(im, (kernel_size, kernel_size))
now_bbox = []
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
now_bbox.append(pt)
j += 1
x, y, w, h = pt[0], pt[1], pt[2] - pt[0], pt[3] - pt[1]
mask = generate_mask(im.shape[0], im.shape[1],
max(w, h) / 2, x + w / 2, y + h / 2)
mask_img[mask] = [0, 0, 0]
add_bbox = find_no_match_box(pre_bbox, now_bbox)
for bbox in add_bbox:
score = bbox[0, i, j, 0]
pt = bbox[0, i, j, 1:] * scale.numpy()
j += 1
x, y, w, h = pt[0], pt[1], pt[2] - pt[0], pt[3] - pt[1]
mask = generate_mask(im.shape[0], im.shape[1],
max(w, h) / 2, x + w / 2, y + h / 2)
mask_img[mask] = [0, 0, 0]
now_bbox += add_bbox
mask_img_verse = np.ones(img.shape, np.int8) - mask_img
result_img = mask_img * im + mask_img_verse * blur_img
#t2 = time.time()
#print('detect:{} timer:{}'.format(img_path, t2 - t1))
#cv2.imwrite(os.path.join(args.save_dir, os.path.basename(img_path)), result_img)
return result_img, now_bbox
def detect(net, img_path, thresh,use_cuda):
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
#max_im_shrink = np.sqrt(1400 * 1200 / (img.shape[0] * img.shape[1]))
max_im_shrink=1.8
image = cv2.resize(img, None, None, fx=max_im_shrink,fy=max_im_shrink, interpolation=cv2.INTER_LINEAR) #这种是对原图像width,height进行缩放,fx,fy表示x和y轴的缩放比例
#image = cv2.resize(img, (800, 600))
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor([img.shape[1], img.shape[0],
img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
img1=img.copy()
left=[]
right=[]
score=[]
for i in range(detections.size(1)):
j = 0
#print(detections[0, i, j, 0])
while detections[0, i, j, 0] >= 0.8:
score.append((detections[0, i, j, 0],j))
j += 1
if len(score)>0:
score=sorted(score,key=lambda value:value[0],reverse=True) #以list中tuple的第一个元素进行降序排列
#sorted()函数不改变原list,sort()函数会改变,默认升序,reverse=True实现降序
num=0
for score in score:
if num>1: #检测到的包含手的概率最高的两个框
break
index=int(score[1])
score=score[0]
pt = (detections[0, i, index, 1:] * scale).cpu().numpy()
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
#j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2) #左上和右下角坐标
left.append(left_up)
right.append(right_bottom)
conf = "{:.3f}".format(score) #保留3位小数
point = (int(left_up[0]), int(left_up[1] - 5))
cv2.putText(img, conf, point, cv2.FONT_HERSHEY_COMPLEX,0.6, (0, 255, 0), 1)
num+=1
#else:
#continue #对于for..else循环,执行完里面的for循环后再执行else,continue结束当前for循环(外面的),执行下次for循环
#break
hand_left=None
hand_right=None
#得到左手的框
if (len(left)==2):
if left[0][0]<left[1][0]:
left_point=left[0]
else:
left_point=left[1]
for i in range(len(left)):
if left_point==left[i]:
left_axis=i
else:
right_axis=i
left_width=right[left_axis][0]-left_point[0]
left_height=right[left_axis][1]-left_point[1]
#print('坐标是({},{})'.format(left_width,left_height))
#左上坐标
if(int(left_point[0]-0.3*left_width)<0):
new_left_x=0
else:
new_left_x=int(left_point[0]-0.3*left_width)
if(int(left_point[1]-0.3*left_height)<0):
new_left_y=0
else:
new_left_y=int(left_point[1]-0.3*left_height)
new_left_point=(new_left_x,new_left_y)
#左下坐标
#一开始左手和右手的框就有重叠,右边框就不扩充了
if not(int(right[left_axis][0]>int(left[right_axis][0]))):
if(int(right[left_axis][0]+0.3*left_width)>int(left[right_axis][0])):
new_left_x1=int(left[right_axis][0])
else:
new_left_x1=int(right[left_axis][0]+0.3*left_width)
else:
new_left_x1=int(right[left_axis][0])
if(int(right[left_axis][1]+0.3*left_height)>height):
new_left_y1=height
else:
new_left_y1=int(right[left_axis][1]+0.3*left_height)
new_left_point1=(new_left_x1,new_left_y1)
#得到右手的框
right_point=left[right_axis] #左上角坐标
right_width=right[right_axis][0]-right_point[0]
right_height=right[right_axis][1]-right_point[1]
#print('坐标是({},{})'.format(right_width,right_height))
#一开始左手和右手的框就有重叠,左边框就不扩充了
if not(int(right_point[0])<int(right[left_axis][0])):
if(int(right_point[0]-0.3*right_width)<int(right[left_axis][0])):
new_right_x=int(right[left_axis][0])
else:
new_right_x=int(right_point[0]-0.3*right_width)
else:
new_right_x=int(right_point[0])
if(int(right_point[1]-0.4*right_height)<0):
new_right_y=0
else:
new_right_y=int(right_point[1]-0.4*right_height)
new_right_point=(new_right_x,new_right_y)
#右下坐标
if(int(right[right_axis][0]+0.3*right_width)>width):
new_right_x1=width
else:
new_right_x1=int(right[right_axis][0]+0.3*right_width)
if(int(right[right_axis][1]+0.4*right_height)>height):
new_right_y1=height
else:
new_right_y1=int(right[right_axis][1]+0.4*right_height)
new_right_point1=(new_right_x1,new_right_y1)
box1=(new_left_point,new_left_point1)
box2=(new_right_point,new_right_point1)
t2 = time.time()
#print('detect:{} timer:{}'.format(img_path, t2 - t1))
#cropImg_right=img1[new_right_point[1]:new_right_point1[1],new_right_point[0]:new_right_point1[0]]
#cv2.imwrite(os.path.join('.', os.path.basename(img_path)), cropImg_right)
return box1,box2
elif len(left==1):
def saveTracesNpy(net, img_list):
if os.path.isdir(args.save_traces_npy_dir) is False:
try:
os.mkdir(args.save_traces_npy_dir)
except OSError:
print("Creation of the directory %s failed" %
args.save_traces_npy_dir)
return
if os.path.isdir(os.path.join(args.save_traces_npy_dir,
'inputs')) is False:
try:
os.mkdir(os.path.join(args.save_traces_npy_dir, 'inputs'))
except OSError:
print("Creation of the directory %s failed" %
os.path.join(args.save_traces_npy_dir, 'inputs'))
return
if os.path.isdir(os.path.join(args.save_traces_npy_dir,
'outputs')) is False:
try:
os.mkdir(os.path.join(args.save_traces_npy_dir, 'outputs'))
except OSError:
print("Creation of the directory %s failed" %
os.path.join(args.save_traces_npy_dir, 'outputs'))
return
inputDims = net.rnn_model.inputDims
nRows = net.rnn_model.nRows
nCols = net.rnn_model.nCols
count = 0
for img_path in img_list:
img = Image.open(os.path.join(args.image_folder, img_path))
img = img.convert('RGB')
img = np.array(img)
max_im_shrink = np.sqrt(640 * 480 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
patches = activation['prepatch']
patches = torch.cat(torch.unbind(patches, dim=2), dim=0)
patches = torch.reshape(patches, (-1, inputDims, nRows, nCols))
rnnX = activation['rnn_model']
patches_all = torch.stack(torch.split(patches,
split_size_or_sections=1,
dim=0),
dim=-1)
rnnX_all = torch.stack(torch.split(rnnX,
split_size_or_sections=1,
dim=0),
dim=-1)
for k in range(patches_all.shape[-1]):
patches_tosave = patches_all[0, :, :, :,
k].cpu().numpy().transpose(1, 2, 0)
rnnX_tosave = rnnX_all[0, :, k].cpu().numpy()
np.save(
args.save_traces_npy_dir + '/inputs/trace_' + str(count) +
'_' + str(k) + '.npy', patches_tosave)
np.save(
args.save_traces_npy_dir + '/outputs/trace_' + str(count) +
'_' + str(k) + '.npy', rnnX_tosave)
count += 1
def detect_hand(net, img_path, thresh, use_cuda, save_path):
if not os.path.exists(save_path):
os.mkdir(save_path)
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img) #--np.array()类型才有shape函数使用
height, width, _ = img.shape
Rect = (32, 499, 1260, 639)
img = process_img(img, Rect) #---将钢琴键盘上方的像素处理一下,
#print(height,width)
#max_im_shrink = np.sqrt(1700 * 1200 / (img.shape[0] * img.shape[1]))
max_im_shrink = 1.5
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR
) #这种是对原图像width,height进行缩放,fx,fy表示x和y轴的缩放比例
#image = cv2.resize(img, (640, 640))
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
img1 = img.copy()
left = []
right = []
scores = []
h_type = []
num_class = []
num_class.append('left')
num_class.append('right')
#i=0的时候检测到的是background class
#i=1是左手,2是右手
for i in range(1, detections.size(1)):
j = 0
print('the probability of {} is {}'.format(num_class[i - 1],
detections[0, i, j, 0]))
#detections[0, i, j, 0]为检测到的手的box的概率,j表示有多少个框(box)
while detections[0, i, j, 0] >= 0.2:
scores.append((detections[0, i, j, 0], j))
j += 1
#print(scores)
if len(scores) > 0:
scores = sorted(scores, key=lambda value: value[0],
reverse=True) #以list中tuple的第一个元素进行降序排列
#sorted()函数不改变原list,sort()函数会改变,默认升序,reverse=True实现降序
num = 0
for score in scores:
if num > 0: #检测到的包含左/右手的概率最高的那个框
break
index = int(score[1])
score = score[0]
pt = (detections[0, i, index, 1:] * scale).cpu().numpy()
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
#j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255),
2) #左上和右下角坐标
left.append(left_up)
right.append(right_bottom)
h_type.append(num_class[i - 1])
conf = "{:.3f}".format(score) #保留3位小数
point = (int(left_up[0]), int(left_up[1] - 5))
point1 = (int(right_bottom[0] - 20), int(left_up[1] - 5))
cv2.putText(img, conf, point, cv2.FONT_HERSHEY_COMPLEX, 0.6,
(0, 0, 255), 2)
cv2.putText(img, str(num_class[i - 1]), point1,
cv2.FONT_HERSHEY_COMPLEX, 0.7, (0, 0, 255), 2)
num += 1
#cv2.imwrite(os.path.join('/home/lj/cy/openpose/piano/test_piano/image/point_dir',os.path.basename(img_path)),img)
#else:
#continue #对于for..else循环,执行完里面的for循环后再执行else,continue结束当前for循环(外面的),执行下次for循环
#break
#双手覆盖的问题先不考虑,比较复杂,看其他的
hand_left = None
hand_right = None
#h_type = None
#再考虑一个手的情况兄弟
if (len(left) == 2):
#下面是假设左手一直在左边,在右手的左边,没有考虑左手在右边(两手交叉),框扩充的时候需要变一下
#而且这种两个手是假定弹钢琴是横着的情况,如果是竖着的话坐标轴变了,判断两个框是否重合的时候进行比较是不一样的
for i, left1 in enumerate(left):
hand_type = h_type[i]
if (hand_type == 'left'):
left_width = right[i][0] - left1[0]
left_height = right[i][1] - left1[1]
#左手框左上坐标(将检测到的框进行扩充,尽量使得手在中心位置,便于检测关键点) ,而且有时候检测到的手不完整
if (int(left1[0] - 0.3 * left_width) < 0):
new_left_x = 0
else:
new_left_x = int(left1[0] - 0.3 * left_width)
if (int(left1[1] - 0.3 * left_height) < 0):
new_left_y = 0
else:
new_left_y = int(left1[1] - 0.3 * left_height)
new_left_point = (new_left_x, new_left_y)
#左手框右下坐标
#一开始左手和右手的框就有重叠,右边框就不扩充了,以防止框中包含两个手
if (i == 0):
j = 1
else:
j = 0
if not (int(right[i][0] > int(left[j][0]))): #左手框右下大于右手框左上
if (int(right[i][0] + 0.3 * left_width) > int(left[j][0])):
new_left_x1 = int(left[j][0]) #如果没重叠左手框右边界最多扩充到右手框的左边界
else:
new_left_x1 = int(right[i][0] + 0.3 * left_width)
else:
new_left_x1 = int(right[i][0])
if (int(right[i][1] + 0.3 * left_height) > height):
new_left_y1 = height
else:
new_left_y1 = int(right[i][1] + 0.3 * left_height)
new_left_point1 = (new_left_x1, new_left_y1)
if (hand_type == 'right'):
#right_point=left[i] #左上角坐标
right_width = right[i][0] - left1[0]
right_height = right[i][1] - left1[1]
if (i == 0):
j = 1
else:
j = 0
#右手框左上坐标
#一开始左手和右手的框就有重叠,左边框就不扩充了
if not (int(left1[0]) < int(right[j][0])):
if (int(left1[0] - 0.3 * right_width) < int(right[j][0])):
new_right_x = int(right[j][0])
else:
new_right_x = int(left1[0] - 0.3 * right_width)
else:
new_right_x = int(left1[0])
if (int(left1[1] - 0.4 * right_height) < 0):
new_right_y = 0
else:
new_right_y = int(left1[1] - 0.4 * right_height)
new_right_point = (new_right_x, new_right_y)
#右手框右上坐标
#右下坐标
if (int(right[i][0] + 0.3 * right_width) > width):
new_right_x1 = width
else:
new_right_x1 = int(right[i][0] + 0.3 * right_width)
if (int(right[i][1] + 0.4 * right_height) > height):
new_right_y1 = height
else:
new_right_y1 = int(right[i][1] + 0.4 * right_height)
new_right_point1 = (new_right_x1, new_right_y1)
box1 = (new_left_point, new_left_point1) #左上和右下角坐标
box2 = (new_right_point, new_right_point1)
cv2.rectangle(img, box1[0], box1[1], (0, 0, 255), 2)
cv2.rectangle(img, box2[0], box2[1], (0, 0, 255), 2)
cv2.imwrite(os.path.join(save_path, os.path.basename(img_path)), img)
return box1, box2, h_type
elif (len(left) == 1):
hand_type = h_type[0]
if (hand_type == 'left'):
left_width = right[0][0] - left[0][0]
left_height = right[0][1] - left[0][1]
#左手框左上坐标
if (int(left[0][0] - 0.3 * left_width) < 0):
new_left_x = 0
else:
new_left_x = int(left[0][0] - 0.3 * left_width)
if (int(left[0][1] - 0.3 * left_height) < 0):
new_left_y = 0
else:
new_left_y = int(left[0][1] - 0.3 * left_height)
new_left_point = (new_left_x, new_left_y)
#左手框右下坐标
if (int(right[0][0] + 0.3 * left_width) > width):
new_left_x1 = width
else:
new_left_x1 = int(right[0][0] + 0.3 * left_width)
if (int(right[0][1] + 0.3 * left_height) > height):
new_left_y1 = height
else:
new_left_y1 = int(right[0][1] + 0.3 * left_height)
new_left_point1 = (new_left_x1, new_left_y1)
box1 = (new_left_point, new_left_point1)
cv2.rectangle(img, box1[0], box1[1], (0, 0, 255), 2)
cv2.imwrite(os.path.join(save_path, os.path.basename(img_path)),
img)
return box1, hand_right, h_type
else:
right_width = right[0][0] - left[0][0]
right_height = right[0][1] - left[0][1]
#右手框左上坐标
if (int(left[0][0] - 0.3 * right_width) < 0):
new_right_x = 0
else:
new_right_x = int(left[0][0] - 0.3 * right_width)
if (int(left[0][1] - 0.3 * right_height) < 0):
new_right_y = 0
else:
new_right_y = int(left[0][1] - 0.3 * right_height)
new_right_point = (new_right_x, new_right_y)
#右手框右下坐标
if (int(right[0][0] + 0.3 * right_width) > width):
new_right_x1 = width
else:
new_right_x1 = int(right[0][0] + 0.3 * right_width)
if (int(right[0][1] + 0.3 * right_height) > height):
new_right_y1 = height
else:
new_right_y1 = int(right[0][1] + 0.3 * right_height)
new_right_point1 = (new_right_x1, new_right_y1)
box2 = (new_right_point, new_right_point1)
cv2.rectangle(img, box2[0], box2[1], (0, 0, 255), 2)
cv2.imwrite(os.path.join(save_path, os.path.basename(img_path)),
img)
return hand_left, box2, h_type
else:
return hand_left, hand_right, h_type
def detect(net, img_path, thresh, gt_path):
img_name = img_path.split('/')[-1].split('.')[0]
out_f = "{}/txt/{}.txt".format(args.save_dir, img_name)
# if os.path.isfile(out_f):
# print("exists.")
# return
fout = open(out_f, "w")
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1500 * 1000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
# plot ground truth
gt = load_gt(gt_path)
for box in gt:
left_up = (box[0], box[1])
right_bottom = (box[2], box[3])
cv2.rectangle(img, left_up, right_bottom, (0, 255, 78), 1)
for i in range(detections.size(1)):
j = 0
while j < detections.size(2) and detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 1)
conf = "{:.2f}".format(score)
text_size, baseline = cv2.getTextSize(conf,
cv2.FONT_HERSHEY_SIMPLEX,
0.3, 1)
p1 = (left_up[0], left_up[1] - text_size[1])
fout.write('{} {} {} {} {}\n'.format(pt[0], pt[1], pt[2], pt[3],
score))
cv2.rectangle(img, (p1[0] - 2 // 2, p1[1] - 2 - baseline),
(p1[0] + text_size[0], p1[1] + text_size[1]),
[255, 0, 0], -1)
cv2.putText(img, conf, (p1[0], p1[1] + baseline),
cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 255, 255), 1, 8)
t2 = time.time()
print('detect:{} timer:{}'.format(img_path, t2 - t1))
cv2.imwrite(os.path.join(args.save_dir, "png", os.path.basename(img_path)),
img)
fout.close()
def detect(net, img_path, thresh, imgName):
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1500 * 1000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= np.array([103.939, 116.779, 123.68])[:, np.newaxis,
np.newaxis].astype('float32')
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
# print(detections)
# print(detections.size())
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
f = open(os.path.join(args.res_dir, imgName.replace('png', 'txt')), "w")
for i in range(detections.size(1)):
j = 0
# while detections[0, i, j, 0] >= thresh:
while ((j < detections.size(2)) and detections[0, i, j, 0] > thresh):
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
score = detections[0, i, j, 0]
f.write("%f %f %f %f %f\n" % (pt[0], pt[1], pt[2], pt[3], score))
j += 1
f.close()
"""
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
conf = "{:.2f}".format(score)
text_size, baseline = cv2.getTextSize(
conf, cv2.FONT_HERSHEY_SIMPLEX, 0.3, 1)
p1 = (left_up[0], left_up[1] - text_size[1])
cv2.rectangle(img, (p1[0] - 2 // 2, p1[1] - 2 - baseline),
(p1[0] + text_size[0], p1[1] + text_size[1]),[255,0,0], -1)
cv2.putText(img, conf, (p1[0], p1[
1] + baseline), cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 255, 255), 1, 8)
"""
t2 = time.time()
print('detect:{} timer:{}'.format(img_path, t2 - t1))
def test_net(save_folder, net, dataset, thresh=0.05):
num_images = len(dataset)
all_boxes = [[[] for _ in range(num_images)] for _ in range(2)]
_t = {'im_detect': Timer(), 'misc': Timer()}
output_dir = get_output_dir(os.path.join(save_folder, 'sfd_hand'),
set_type)
det_file = os.path.join(output_dir, 'detections.pkl')
for i in range(num_images):
img = dataset.pull_image(i)
h, w, _ = img.shape
shrink = np.sqrt(1700 * 1200 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=shrink,
fy=shrink,
interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
_t['im_detect'].tic()
detections = net(x).data
detect_time = _t['im_detect'].toc(average=False)
for j in range(1, detections.size(1)):
dets = detections[0, j, :]
mask = dets[:, 0].gt(thresh).expand(5, dets.size(0)).t()
dets = torch.masked_select(dets, mask).view(-1, 5)
if dets.dim() == 0:
continue
boxes = dets[:, 1:]
boxes[:, 0] *= w
boxes[:, 2] *= w
boxes[:, 1] *= h
boxes[:, 3] *= h
scores = dets[:, 0].cpu().numpy()
cls_dets = np.hstack(
(boxes.cpu().numpy(), scores[:,
np.newaxis])).astype(np.float32,
copy=False)
all_boxes[j][i] = cls_dets
fin_mask = np.where(scores > 0.6)[0]
bboxes = boxes.cpu().numpy()[fin_mask]
scores = scores[fin_mask]
for k in range(len(scores)):
leftup = (int(bboxes[k][0]), int(bboxes[k][1]))
right_bottom = (int(bboxes[k][2]), int(bboxes[k][3]))
cv2.rectangle(img, leftup, right_bottom, (0, 255, 0), 2)
save_file = os.path.join(output_dir, '{}.jpg'.format(i + 1))
cv2.imwrite(save_file, img)
print('im_detect: {:d}/{:d} {:.3f}s'.format(i + 1, num_images,
detect_time))
with open(det_file, 'wb') as f:
pickle.dump(all_boxes, f, pickle.HIGHEST_PROTOCOL)
print('Evaluating detections')
evaluate_detections(all_boxes, output_dir, dataset)
def detect(net, img_path, thresh, use_cuda):
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
print(height, width)
#max_im_shrink = np.sqrt(1700 * 1200 / (img.shape[0] * img.shape[1]))
max_im_shrink = 1.5
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR
) #这种是对原图像width,height进行缩放,fx,fy表示x和y轴的缩放比例
#image = cv2.resize(img, (640, 640))
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
img1 = img.copy()
left = []
right = []
scores = []
h_type = []
num_class = []
num_class.append('left')
num_class.append('right')
#i=0的时候检测到的是background class
#i=1是左手,2是右手
for i in range(1, detections.size(1)):
j = 0
print('the pro of {} is {}'.format(num_class[i - 1], detections[0, i,
j, 0]))
#detections[0, i, j, 0]为检测到的手的box的概率,j表示有多少个框(box)
while detections[0, i, j, 0] >= 0.7:
scores.append((detections[0, i, j, 0], j))
j += 1
#print(scores)
if len(scores) > 0:
scores = sorted(scores, key=lambda value: value[0],
reverse=True) #以list中tuple的第一个元素进行降序排列
#sorted()函数不改变原list,sort()函数会改变,默认升序,reverse=True实现降序
num = 0
for t_index, score in enumerate(scores):
if num > 0: #检测到的包含手的概率最高的两个框,因为有时候会将一些没有包含手的信息误认为手从而检测
break
index = int(score[1])
score = score[0]
pt = (detections[0, i, index, 1:] * scale).cpu().numpy()
left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
#j += 1
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255),
2) #左上和右下角坐标
left.append(left_up)
right.append(right_bottom)
conf = "{:.3f}".format(score) #保留3位小数
point = (int(left_up[0]), int(left_up[1] - 5))
point1 = (int(right_bottom[0] - 20), int(left_up[1] - 5))
cv2.putText(img, conf, point, cv2.FONT_HERSHEY_COMPLEX, 0.6,
(0, 0, 255), 2)
cv2.putText(img, str(num_class[i - 1]), point1,
cv2.FONT_HERSHEY_COMPLEX, 0.7, (0, 0, 255), 2)
num += 1
cv2.imwrite(
os.path.join(
'/home/lj/cy/openpose/piano/test_piano/image/point_dir',
os.path.basename(img_path)), img)
#else:
#continue #对于for..else循环,执行完里面的for循环后再执行else,continue结束当前for循环(外面的),执行下次for循环
#break
hand_left = None
hand_right = None
#(检测到两只手时,根据框的位置判断是左手还是右手),如果两手交叉是不是会出现问题
if (len(left) == 2):
if left[0][0] < left[1][0]:
left_point = left[0]
else:
left_point = left[1]
for i in range(len(left)):
if left_point == left[i]:
left_axis = i
else:
right_axis = i
left_width = right[left_axis][0] - left_point[0]
left_height = right[left_axis][1] - left_point[1]
#print('坐标是({},{})'.format(left_width,left_height))
#左上坐标(将检测到的框进行扩充,尽量使得手在中心位置,便于检测关键点)
if (int(left_point[0] - 0.3 * left_width) < 0):
new_left_x = 0
else:
new_left_x = int(left_point[0] - 0.3 * left_width)
if (int(left_point[1] - 0.3 * left_height) < 0):
new_left_y = 0
else:
new_left_y = int(left_point[1] - 0.3 * left_height)
new_left_point = (new_left_x, new_left_y)
#左下坐标
#一开始左手和右手的框就有重叠,右边框就不扩充了
if not (int(right[left_axis][0] > int(left[right_axis][0]))):
if (int(right[left_axis][0] + 0.3 * left_width) > int(
left[right_axis][0])):
new_left_x1 = int(left[right_axis][0])
else:
new_left_x1 = int(right[left_axis][0] + 0.3 * left_width)
else:
new_left_x1 = int(right[left_axis][0])
if (int(right[left_axis][1] + 0.3 * left_height) > height):
new_left_y1 = height
else:
new_left_y1 = int(right[left_axis][1] + 0.3 * left_height)
new_left_point1 = (new_left_x1, new_left_y1)
#得到右手的框
right_point = left[right_axis] #左上角坐标
right_width = right[right_axis][0] - right_point[0]
right_height = right[right_axis][1] - right_point[1]
#print('坐标是({},{})'.format(right_width,right_height))
#一开始左手和右手的框就有重叠,左边框就不扩充了
if not (int(right_point[0]) < int(right[left_axis][0])):
if (int(right_point[0] - 0.3 * right_width) < int(
right[left_axis][0])):
new_right_x = int(right[left_axis][0])
else:
new_right_x = int(right_point[0] - 0.3 * right_width)
else:
new_right_x = int(right_point[0])
if (int(right_point[1] - 0.4 * right_height) < 0):
new_right_y = 0
else:
new_right_y = int(right_point[1] - 0.4 * right_height)
new_right_point = (new_right_x, new_right_y)
#右下坐标
if (int(right[right_axis][0] + 0.3 * right_width) > width):
new_right_x1 = width
else:
new_right_x1 = int(right[right_axis][0] + 0.3 * right_width)
if (int(right[right_axis][1] + 0.4 * right_height) > height):
new_right_y1 = height
else:
new_right_y1 = int(right[right_axis][1] + 0.4 * right_height)
new_right_point1 = (new_right_x1, new_right_y1)
box1 = (new_left_point, new_left_point1) #左上和右下角坐标
box2 = (new_right_point, new_right_point1)
t2 = time.time()
#print('detect:{} timer:{}'.format(img_path, t2 - t1))
#cropImg_right=img1[new_right_point[1]:new_right_point1[1],new_right_point[0]:new_right_point1[0]]
#cv2.imwrite(os.path.join('.', os.path.basename(img_path)), cropImg_right)
return box1, box2
else:
return hand_left, hand_right
def detect(net, img_path, thresh):
#img = cv2.imread(img_path, cv2.IMREAD_COLOR)
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1700 * 1200 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
#image = cv2.resize(img, (640, 640))
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0), requires_grad=False)
print(x.shape)
if use_cuda:
x = x.cuda()
t1 = time.time()
y = net(x)
detections = y.data #torch.size:[1,2,750,5]
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1],
img.shape[0]]) #original img shape
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
# for i in range(detections.size(1)):
# j = 0
# while detections[0, i, j, 0] >= thresh: #5iterms the first is the score
# score = detections[0, i, j, 0]
# pt = (detections[0, i, j, 1:] * scale).cpu().numpy() #left up x,left up y ,right bottom x ,right bottom y
# left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
# j += 1
# cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
# conf = "{:.3f}".format(score)
# point = (int(left_up[0]), int(left_up[1] - 5))
# #cv2.putText(img, conf, point, cv2.FONT_HERSHEY_COMPLEX,
# # 0.6, (0, 255, 0), 1)
dclone = []
for i in range(detections.size(1)):
j = 0
while detections[0, i, j,
0] >= thresh: #5iterms the first is the score
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy(
) #left up x,left up y ,right bottom x ,right bottom y
dclone.append([pt[0], pt[1], pt[2], pt[3], score])
j += 1
dclone = np.array(dclone)
dclone = dclone[[my_nms(np.array(dclone), 0.3)]]
for i in range(dclone.shape[0]):
pt = dclone[i]
left_up, right_bottom = (int(pt[0]), int(pt[1])), (int(pt[2]),
int(pt[3]))
cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
t2 = time.time()
print('detect:{} timer:{}'.format(img_path, t2 - t1))
cv2.imwrite(os.path.join(args.save_dir, os.path.basename(img_path)), img)
def detect(net, im, thresh):
#img = Image.open(img_path)
img = Image.fromarray(cv2.cvtColor(im, cv2.COLOR_BGR2RGB))
if img.mode == 'L':
img = img.convert('RGB')
img = np.array(img)
height, width, _ = img.shape
max_im_shrink = np.sqrt(1200 * 1000 / (img.shape[0] * img.shape[1]))
image = cv2.resize(img,
None,
None,
fx=max_im_shrink,
fy=max_im_shrink,
interpolation=cv2.INTER_LINEAR)
#image = cv2.resize(img,(224,),interpolation=cv2.INTER_LINEAR)
x = to_chw_bgr(image)
x = x.astype('float32')
x -= cfg.img_mean
x = x[[2, 1, 0], :, :]
x = Variable(torch.from_numpy(x).unsqueeze(0))
if use_cuda:
x = x.cuda()
#t1 = time.time()
y = net(x)
detections = y.data
scale = torch.Tensor(
[img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
#img = cv2.imread(img_path, cv2.IMREAD_COLOR)
mask_img = np.ones(im.shape, np.int8)
kernel_size = 15
blur_img = cv2.blur(im, (kernel_size, kernel_size))
for i in range(detections.size(1)):
j = 0
while detections[0, i, j, 0] >= thresh:
score = detections[0, i, j, 0]
pt = (detections[0, i, j, 1:] * scale).cpu().numpy().astype(int)
#left_up, right_bottom = (pt[0], pt[1]), (pt[2], pt[3])
j += 1
#cv2.rectangle(img, left_up, right_bottom, (0, 0, 255), 2)
#conf = "{:.2f}".format(score)
#text_size, baseline = cv2.getTextSize(
# conf, cv2.FONT_HERSHEY_SIMPLEX, 0.3, 1)
#p1 = (left_up[0], left_up[1] - text_size[1])
#cv2.rectangle(img, (p1[0] - 2 // 2, p1[1] - 2 - baseline),
# (p1[0] + text_size[0], p1[1] + text_size[1]),[255,0,0], -1)
#cv2.putText(img, conf, (p1[0], p1[
# 1] + baseline), cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 255, 255), 1, 8)
x, y, w, h = pt[0], pt[1], pt[2] - pt[0], pt[3] - pt[1]
mask = generate_mask(im.shape[0], im.shape[1],
max(w, h) / 2, x + w / 2, y + h / 2)
mask_img[mask] = [0, 0, 0]
mask_img_verse = np.ones(img.shape, np.int8) - mask_img
result_img = mask_img * im + mask_img_verse * blur_img
#t2 = time.time()
#print('detect:{} timer:{}'.format(img_path, t2 - t1))
#cv2.imwrite(os.path.join(args.save_dir, os.path.basename(img_path)), result_img)
return result_img
