def extract_features(model,
data_loader,
is_flip=False,
print_freq=1,
metric=None):
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
features = OrderedDict()
end = time.time()
if is_flip:
print('flip')
for i, (imgs, flip_imgs, fnames) in enumerate(data_loader):
data_time.update(time.time() - end)
outputs = extract_cnn_feature(model, imgs)
flip_outputs = extract_cnn_feature(model, flip_imgs)
final_outputs = (outputs + flip_outputs) / 2
for fname, output in zip(fnames, final_outputs):
features[fname] = output.numpy()
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % print_freq == 0:
print('Extract Features: [{}/{}]\t'
'Time {:.3f} ({:.3f})\t'
'Data {:.3f} ({:.3f})\t'.format(i + 1, len(data_loader),
batch_time.val,
batch_time.avg,
data_time.val,
data_time.avg))
else:
print('no flip')
for i, (imgs, fnames) in enumerate(data_loader):
data_time.update(time.time() - end)
outputs = extract_cnn_feature(model, imgs)
for fname, output in zip(fnames, outputs):
features[fname] = output.numpy()
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % print_freq == 0:
print('Extract Features: [{}/{}]\t'
'Time {:.3f} ({:.3f})\t'
'Data {:.3f} ({:.3f})\t'.format(i + 1, len(data_loader),
batch_time.val,
batch_time.avg,
data_time.val,
data_time.avg))
return features
def reid_draw(frame, b_b, model, shujuku, frame_size):
size = frame_size
left=int((b_b[0] - b_b[2]/2.0) * size[0])
top=int((b_b[1]- b_b[3]/2.0) * size[1])
right=int((b_b[0] + b_b[2]/2.0) * size[0])
bottom=int((b_b[1] + b_b[3]/2.0) * size[1])
if top>=bottom or left>=right or top<=0 or left<=0:
return
img1 = jieduan(frame,left,top,right,bottom)
img = preprocess(img1)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
#for feature2,filename in shujuku:
customer_id = -1
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature,fea)
if minsim > distan or minsim == -1:
minsim = distan
customer_id = int(query)
#add to debug customer_id
#customer_id = 1
cv2.rectangle(frame, (left, top), (right, bottom), (255, 0, 0), 2)
cv2.putText(frame,str(customer_id),(left,top),cv2.FONT_HERSHEY_COMPLEX,6,(255,0,0),2)
customer_name = "name_"+str(customer_id)
assert type(customer_id) is int # must be number
# return frame
#print("pub msg finish")
return customer_name, customer_id
def extract_features(model, data_loader, print_freq=50, metric=None):
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
features = OrderedDict()
labels = OrderedDict()
end = time.time()
for i, (imgs, fnames, pids, _) in enumerate(data_loader):
data_time.update(time.time() - end)
outputs = extract_cnn_feature(model, imgs)
for fname, output, pid in zip(fnames, outputs, pids):
features[fname] = output
labels[fname] = pid
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % print_freq == 0:
print('Extract Features: [{}/{}]\t'
'Time {:.3f} ({:.3f})\t'
'Data {:.3f} ({:.3f})\t'.format(i + 1, len(data_loader),
batch_time.val,
batch_time.avg,
data_time.val,
data_time.avg))
return features, labels
def extract_features_cross_cam(model, data_loader):
model.eval()
cross_cam_features = []
cross_cam_fnames = []
cross_cam_distribute = []
cams = []
cam_number = len(model.classifier)
print("Start extract features cross camera")
for imgs, fnames, _, camid in tqdm(data_loader):
with torch.no_grad():
outputs = extract_cnn_feature(model, imgs, norm=False)
fnorm = torch.norm(outputs, p=2, dim=1, keepdim=True)
norm_outputs = outputs.div(fnorm.expand_as(outputs))
for i in range(cam_number):
x = model.classifier[i](outputs)
if i == 0:
distribute = F.softmax(x.data, dim=1)
else:
distribute_tmp = F.softmax(x.data, dim=1)
distribute = torch.cat((distribute, distribute_tmp), dim=1)
for fname, output, cam, dis in zip(fnames, norm_outputs, camid,
distribute):
cam = cam.item()
cross_cam_fnames.append(fname)
cross_cam_features.append(output)
cams.append(cam)
cross_cam_distribute.append(dis.cpu().numpy())
return cross_cam_features, cross_cam_fnames, cross_cam_distribute, cams
def extract_features(model, data_loader, eval_only, print_freq=100):
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
features = []
labels = []
cameras = []
end = time.time()
for i, (imgs, fnames, pids, cids) in enumerate(data_loader):
data_time.update(time.time() - end)
outputs = extract_cnn_feature(model, imgs, eval_only)
for fname, output, pid, cid in zip(fnames, outputs, pids, cids):
features.append(output)
labels.append(int(pid.numpy()))
cameras.append(int(cid.numpy()))
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % print_freq == 0:
print('Extract Features: [{}/{}]\t'
'Time {:.3f} ({:.3f})\t'
'Data {:.3f} ({:.3f})\t'
.format(i + 1, len(data_loader),
batch_time.val, batch_time.avg,
data_time.val, data_time.avg))
output_features = torch.stack(features, 0)
return output_features, labels, cameras
def reid_draw(frame, b_b, model, cfg):
global size
id_name = 0
cfg.cuda()
left = int((b_b[0] - b_b[2] / 2.0) * size[0])
top = int((b_b[1] - b_b[3] / 2.0) * size[1])
right = int((b_b[0] + b_b[2] / 2.0) * size[0])
bottom = int((b_b[1] + b_b[3] / 2.0) * size[1])
#print("bottom is {}".format(bottom))
if left < 0 or right < 0 or top < 0 or bottom < 0:
return left, top, right, bottom, 999
# 处理半身
if bottom > 530:
ratio = float(bottom - top) / (right - left)
#print("ratio is: {}".format(ratio))
if ratio < 1.5:
#print("ratio is: {}".format(ratio))
#print('filtered out')
return left, top, right, bottom, 999
frame_reid = copy.deepcopy(frame)
# draw shangpin area
left_x, top_y, right_m, bottom_n = shangpin_area()
cv2.rectangle(frame, (left_x, top_y), (right_m, bottom_n), (0, 255, 0), 2)
left_x_2, top_y_2, right_m_2, bottom_n_2 = shangpin_area_huojia2()
cv2.rectangle(frame, (left_x_2, top_y_2), (right_m_2, bottom_n_2),
(255, 0, 0), 2)
img1 = jieduan(frame_reid, left, top, right, bottom)
img = preprocess(img1)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
try:
pkl_file = open('/data/reid/renti/data.pkl', 'rb')
shujuku = pickle.load(pkl_file)
pkl_file.close()
except:
pkl_file = open('/data/reid/renti/data_bu.pkl', 'rb')
shujuku = pickle.load(pkl_file)
pkl_file.close()
# for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature, fea)
if minsim > distan or minsim == -1:
minsim = distan
id_name = int(query)
cv2.rectangle(frame, (left, top), (right, bottom), (255, 0, 0), 2)
cv2.putText(frame, str(id_name), (left, top), cv2.FONT_HERSHEY_COMPLEX, 6,
(255, 0, 0), 2)
return left, top, right, bottom, id_name
def extract_n_save(model,
data_loader,
args,
root,
num_cams,
is_detection=True,
use_fname=True,
gt_type='reid'):
model.eval()
print_freq = 1000
batch_time = AverageMeter()
data_time = AverageMeter()
if_created = [0 for _ in range(num_cams)]
lines = [[] for _ in range(num_cams)]
end = time.time()
for i, (imgs, fnames, pids, cams) in enumerate(data_loader):
cams += 1
outputs = extract_cnn_feature(model, imgs)
for fname, output, pid, cam in zip(fnames, outputs, pids, cams):
if is_detection:
pattern = re.compile(r'c(\d+)_f(\d+)')
cam, frame = map(int, pattern.search(fname).groups())
# f_names[cam - 1].append(fname)
# features[cam - 1].append(output.numpy())
line = np.concatenate(
[np.array([cam, 0, frame]),
output.numpy()])
else:
if use_fname:
pattern = re.compile(r'(\d+)_c(\d+)_f(\d+)')
pid, cam, frame = map(int, pattern.search(fname).groups())
else:
cam, pid = cam.numpy(), pid.numpy()
frame = -1 * np.ones_like(pid)
# line = output.numpy()
line = np.concatenate(
[np.array([cam, pid, frame]),
output.numpy()])
lines[cam - 1].append(line)
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % print_freq == 0:
print('Extract Features: [{}/{}]\t'
'Time {:.3f} ({:.3f})\t'
'Data {:.3f} ({:.3f})\t'.format(i + 1, len(data_loader),
batch_time.val,
batch_time.avg,
data_time.val,
data_time.avg))
if_created = save_file(lines, args, root, if_created)
lines = [[] for _ in range(num_cams)]
save_file(lines, args, root, if_created)
return
def inference(model, query_loader, gallery_loader, use_gpu):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf = []
for batch_idx, (imgs, _) in enumerate(query_loader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = extract_cnn_feature(model, imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
qf.extend(list(features))
gf, g_paths = [], []
for batch_idx, (imgs, path) in enumerate(gallery_loader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = extract_cnn_feature(model, imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.extend(list(features))
g_paths.extend(list(path))
print('=> BatchTime(s): {:.3f}'.format(batch_time.avg))
x = torch.cat([qf[i].unsqueeze(0) for i in range(len(qf))], 0)
y = torch.cat([gf[i].unsqueeze(0) for i in range(len(gf))], 0)
m, n = x.size(0), y.size(0)
x = x.view(m, -1)
y = y.view(n, -1)
dist = torch.pow(x, 2).sum(1).unsqueeze(1).expand(m, n) + \
torch.pow(y, 2).sum(1).unsqueeze(1).expand(n, m).t()
dist.addmm_(1, -2, x, y.t())
return dist
def execute(self, frame: FrameType) -> bytes:
trans_im = cv2.resize(frame, (self.width, self.height))
image = Image.fromarray(trans_im)
img = self.test_transformer(image)
img_list = [img]
imgs = torch.stack(img_list)
img_feat = extract_cnn_feature(self.model, imgs)[0]
output = pickle.dumps(img_feat)
return output
def extract_features(model, data):
model.eval()
features = OrderedDict()
labels = OrderedDict()
uids, pids, imgs = data
outputs = extract_cnn_feature(model, imgs)
for uid, pid, output in zip(uids, pids, outputs):
features[uid] = output
labels[uid] = pid
return features, labels
def execute(self, frame: FrameType) -> FrameType:
trans_im = cv2.resize(frame, (self.width, self.height))
image = Image.fromarray(trans_im)
img = self.test_transformer(image)
img_list = [img]
imgs = torch.stack(img_list)
t1 = datetime.now()
img_feat = extract_cnn_feature(self.model, imgs)[0]
t2 = datetime.now()
# print("extract features in ", (t2 - t1).total_seconds() * 1000, " ms.")
# print(img_feat)
return trans_im
def reid_draw(frame, b_b, model, cfg):
global size
id_name = 'new'
cfg.cuda()
left = int((b_b[0] - b_b[2] / 2.0) * size[0])
top = int((b_b[1] - b_b[3] / 2.0) * size[1])
right = int((b_b[0] + b_b[2] / 2.0) * size[0])
bottom = int((b_b[1] + b_b[3] / 2.0) * size[1])
#print left,top,right,bottom
if left < 0 or right < 0 or top < 0 or bottom < 0:
return left, top, right, bottom, id_name
img1 = jieduan(frame, left, top, right, bottom)
img = preprocess(img1)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
rentidir = '/home/tujh/renti/'
pkl_file = open('/data/reid/renti/data.pkl', 'rb')
shujuku = pickle.load(pkl_file)
#for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature, fea)
if minsim > distan or minsim == -1:
minsim = distan
id_name = int(query)
cv2.rectangle(frame, (left, top), (right, bottom), (255, 0, 0), 2)
cv2.putText(frame, str(id_name), (left, top), cv2.FONT_HERSHEY_COMPLEX, 6,
(255, 0, 0), 2)
# draw shangpin area
left_x, top_y, right_m, bottom_n = shangpin_area()
cv2.rectangle(frame, (left_x, top_y), (right_m, bottom_n), (0, 255, 0), 2)
left_x_2, top_y_2, right_m_2, bottom_n_2 = shangpin_area_huojia2()
cv2.rectangle(frame, (left_x_2, top_y_2), (right_m_2, bottom_n_2),
(255, 0, 0), 2)
#print(left, top, right, bottom)
return left, top, right, bottom, id_name
def extract_features(model,
data_loader,
print_freq=1,
save_name='feature.mat'):
batch_time = AverageMeter()
data_time = AverageMeter()
ids = []
cams = []
features = []
query_files = []
end = time.time()
for i, (imgs, fnames) in enumerate(data_loader):
data_time.update(time.time() - end)
outputs = extract_cnn_feature(model, imgs)
#for test time augmentation
#bs, ncrops, c, h, w = imgs.size()
#outputs = extract_cnn_feature(model, imgs.view(-1,c,h,w))
#outputs = outputs.view(bs,ncrops,-1).mean(1)
for fname, output in zip(fnames, outputs):
if fname[0] == '-':
ids.append(-1)
cams.append(int(fname[4]))
else:
ids.append(int(fname[:4]))
cams.append(int(fname[6]))
features.append(output.numpy())
query_files.append(fname)
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % print_freq == 0:
print('Extract Features: [{}/{}]\t'
'Time {:.3f} ({:.3f})\t'
'Data {:.3f} ({:.3f})\t'.format(i + 1, len(data_loader),
batch_time.val,
batch_time.avg,
data_time.val,
data_time.avg))
return features, ids, cams, query_files
def loadDataset():
torch.cuda.set_device(0)
logs_dir = 'market-1501-Exper33/RPP/'
num_features = 256
num_classes = 751
T = 1
dim = 256
dropout = 0.5
###
model = models.create('resnet50_rpp',
num_features=num_features,
dropout=dropout,
num_classes=num_classes,
cut_at_pooling=False,
FCN=True,
T=T,
dim=dim)
model = model.cuda()
checkpoint = load_checkpoint(osp.join(logs_dir, 'checkpoint.pth.tar'))
model.load_state_dict(checkpoint['state_dict'])
res = []
frame_number = 0
# --datasets
shujuku = {}
rentidir = '/data/reid/renti/queries'
for query in os.listdir(rentidir):
query_dir = osp.join(rentidir, query)
shujuku[query] = []
for img in os.listdir(query_dir):
_img = cv2.imread(osp.join(query_dir, img))
_img = np.transpose(_img, (2, 0, 1)).astype(np.float32)
_img = torch.from_numpy(_img)
_img = torch.unsqueeze(_img, 0)
_feature = extract_cnn_feature(model, _img.cuda())
shujuku[query].append(_feature)
# --
return model, shujuku
def extract_features_per_cam(model, data_loader):
model.eval()
per_cam_features = {}
per_cam_fname = {}
print("Start extract features per camera")
for imgs, fnames, _, camid in tqdm(data_loader):
camid = list(camid)
for cam in camid:
cam = cam.item()
if cam not in per_cam_features.keys():
per_cam_features[cam] = []
per_cam_fname[cam] = []
with torch.no_grad():
outputs = extract_cnn_feature(model, imgs)
for fname, output, cam in zip(fnames, outputs, camid):
cam = cam.item()
per_cam_features[cam].append(output)
per_cam_fname[cam].append(fname)
return per_cam_features, per_cam_fname
def compute_distance(frame, b_b):
left=int((b_b[0] - b_b[2]/2.0) * size[0])
top=int((b_b[1]- b_b[3]/2.0) * size[1])
right=int((b_b[0] + b_b[2]/2.0) * size[0])
bottom=int((b_b[1] + b_b[3]/2.0) * size[1])
img1 = jieduan(frame,left,top,right,bottom)
img = np.transpose(img1, (2,0,1)).astype(np.float32)
img = torch.from_numpy(img)
img = torch.unsqueeze(img, 0)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
#for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature,fea)
if minsim > distan or minsim == -1:
minsim = distan
return minsim
def reid_draw(frame, b_b, model, cfg, huojia1_id, pre_res, change_idnum):
global size
global save_box_no
id_name = 0
cfg.cuda()
left = int((b_b[0] - b_b[2] / 2.0) * size[0])
top = int((b_b[1] - b_b[3] / 2.0) * size[1])
right = int((b_b[0] + b_b[2] / 2.0) * size[0])
bottom = int((b_b[1] + b_b[3] / 2.0) * size[1])
if left < 0 or right < 0 or top < 0 or bottom < 0:
return left, top, right, bottom, 999
# if bottom > 530:
# ratio = float(bottom - top) / (right - left)
# #print("ratio is: {}".format(ratio))
# if ratio < 1.5:
# #print("ratio is: {}".format(ratio))
# print('filtered out')
# return left, top, right, bottom, 999
frame_reid = copy.deepcopy(frame)
# draw shangpin area
left_x, top_y, right_m, bottom_n = shangpin_area(huojia1_id)
cv2.rectangle(frame, (left_x, top_y), (right_m, bottom_n), (0, 255, 0), 2)
ratio = float(bottom - top) / (right - left)
# # print(ratio)
# if ratio < 2.0:
# # print('filtered out')
# return left, top, right, bottom, 999
img1 = jieduan(frame_reid, left, top, right, bottom)
img = preprocess(img1)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
try:
pkl_file = open('/data/reid/renti/data.pkl', 'rb')
shujuku = pickle.load(pkl_file)
pkl_file.close()
except:
pkl_file = open('/data/reid/renti/data_bu.pkl', 'rb')
shujuku = pickle.load(pkl_file)
pkl_file.close()
rentidir = '/home/tujh/renti/'
# pkl_file = open('/data/reid/renti/data.pkl', 'rb')
# shujuku = pickle.load(pkl_file)
# pre_item_huoid={}
# for id_name, pre_item in pre_res.items():
# if huojia1_id == pre_item[-1]:
# pre_item_huoid[id_name]=pre_item ##person id in front of huojia_id
if change_idnum: #len(pre_res) == len(shujuku) and pre_item_huoid:
id_name = reid_draw_multi(pre_res, b_b)
pre_fix = 'B:'
else:
# for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature, fea)
if minsim > distan or minsim == -1:
minsim = distan
id_name = int(query)
pre_fix = 'R:'
cv2.rectangle(frame, (left, top), (right, bottom), (255, 0, 0), 2)
cv2.putText(frame, pre_fix + str(id_name), (left, top),
cv2.FONT_HERSHEY_COMPLEX, 6, (255, 0, 0), 2)
cv2.imwrite(
'/home/zhaocy/yhr/tmp_imgs/' + str(save_box_no) + '_' + str(id_name) +
'.jpg', img1)
save_box_no += 1
return left, top, right, bottom, id_name
def reid_draw(frame, b_b, model, cfg, shujuku, threadPubMsg, camera_id, flag):
global size
# print("size = ", size)
# print("b_b = ", b_b)
# print model
cfg.cuda()
left = int((b_b[0] - b_b[2] / 2.0) * size[0])
top = int((b_b[1] - b_b[3] / 2.0) * size[1])
right = int((b_b[0] + b_b[2] / 2.0) * size[0])
bottom = int((b_b[1] + b_b[3] / 2.0) * size[1])
img1 = jieduan(frame, left, top, right, bottom)
img = np.transpose(img1, (2, 0, 1)).astype(np.float32)
img = torch.from_numpy(img)
img = torch.unsqueeze(img, 0)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
id_name = 'new'
rentidir = '/home/tujh/renti/'
#for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature, fea)
if minsim > distan or minsim == -1:
minsim = distan
id_name = query
cv2.rectangle(frame, (left, top), (right, bottom), (255, 0, 0), 2)
cv2.putText(frame, id_name, (left, top), cv2.FONT_HERSHEY_COMPLEX, 6,
(255, 0, 0), 2)
print(frame.shape)
# left_x = 230
# top_y = 520
# right_m = 480
# bottom_n = 800
# left_x = 89
# top_y = 236
# right_m = 470
# bottom_n = 508
left_x = 600
top_y = 30
right_m = 800
bottom_n = 520
cv2.rectangle(frame, (left_x, top_y), (right_m, bottom_n), (0, 255, 0), 2)
calcIOU1 = calcIOU(left, top, right, bottom, left_x, top_y, right_m,
bottom_n)
if calcIOU1 <= 0 and flag == 0: # back
flag = 0
flag1 = 0
flag2 = 0
elif calcIOU1 > 0 and flag == 0: # in
flag = 1
flag1 = 1
flag2 = 0
elif calcIOU1 <= 0 and flag == 1: # out
flag = 0
flag1 = 0
flag2 = 1
# elif calcIOU1 > 0 and flag == 1: # back
else:
flag = 1
flag1 = 0
flag2 = 0
print(flag)
print(calcIOU1)
print("set customer message")
customer_name = "name" + str(id_name)
customer_id = id_name # number
threadPubMsg.set_customer(name=customer_name,
person_id=12,
camera_id=camera_id,
x=left,
y=top,
w=right,
h=bottom)
threadPubMsg.set_commodity_recognition_trigger(camera_id=camera_id,
flag1=flag1,
flag2=flag2)
# return frame
return flag
def update_shujuku(frame, b_b, new_name, newout):
global print_frame
global x_history
left=int((b_b[0] - b_b[2]/2.0) * size[0])
top=int((b_b[1]- b_b[3]/2.0) * size[1])
right=int((b_b[0] + b_b[2]/2.0) * size[0])
bottom=int((b_b[1] + b_b[3]/2.0) * size[1])
center_x = b_b[0]*size[0]
#print(right- left)
print('bottom:{}'.format(bottom))
print('width:{}'.format(right-left))
print('height:{}'.format(bottom-top))
if top<0 or top>=bottom or left>=right or left<0 or right<0 or bottom<0: #or left<300:
return False
if right - left < 110:
#print('bounding box width too small')
return False
if bottom - top < 150:
#print('bounding box height too small')
return False
#if bottom < 480:
# return False
print('center_x:{}'.format(center_x))
x_history.append(center_x)
print('x_history:')
print(x_history)
img1 = jieduan(frame,left,top,right,bottom)
print_frame += 1
img = preprocess(img1)
#print(left)
#print(top)
#print(right)
#print(bottom)
#print(img.shape)
feature = extract_cnn_feature(model, img.cuda())
if len(shujuku):
minsim = -1
id_name = 'new'
#rentidir = '/home/tujh/renti/'
#for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature,fea)
if minsim > distan or minsim == -1:
minsim = distan
id_name = query
print('minsim:{}'.format(minsim))
print('new_in:' + str(new_in))
new_people_shujuku(feature, new_name)
cv2.imwrite('./tmp2/' + str(print_frame) +'_' + str(new_name) +'.jpg',
img1)
features_at_door.append(feature)
return False
else:
new_people_shujuku(feature, new_name)
features_at_door.append(feature)
cv2.imwrite('./tmp2/' + str(print_frame) +'_' + str(new_name) +'.jpg', img1)
return False
def evaluate():
print('Get dataloader... ')
single_train_loader, single_query_loader, multi_query_loader, gallery_loader = get_dataloader(
)
print('Create and load pre-trained model...')
model = models.create('resnet50',
dropout=0.0,
num_features=2048,
num_classes=632)
# checkpoint = load_checkpoint('./logs/deep-person-1-new-augmentation/market1501-resnet50/model_best.pth.tar')
checkpoint = load_checkpoint('./logs/duke2market/model_best.pth.tar')
model.load_state_dict(checkpoint['state_dict'])
model = nn.DataParallel(model).cuda()
model.eval()
print('Extract train&single_query&gallery feature...')
single_train_feat, single_train_ids, single_train_cams, train_files = extract_features(
model, single_train_loader)
single_query_feat, single_query_ids, single_query_cams, query_files = extract_features(
model, single_query_loader)
gallery_feat, gallery_ids, gallery_cams, _ = extract_features(
model, gallery_loader)
print('Get multi_query feature...')
multi_query_dict = dict()
for i, (imgs, fnames) in enumerate(multi_query_loader):
outputs = extract_cnn_feature(model, imgs)
# test time augmentation
for fname, output in zip(fnames, outputs):
if multi_query_dict.get(fname[:7]) == None:
multi_query_dict[fname[:7]] = []
multi_query_dict[fname[:7]].append(output.numpy())
query_max_feat = []
query_avg_feat = []
for query_file in query_files:
index = query_file[:7]
multi_features = multi_query_dict[index]
multi_features = normalize(multi_features)
query_max_feat.append(np.max(multi_features, axis=0))
query_avg_feat.append(np.mean(multi_features, axis=0))
assert len(query_max_feat) == len(query_avg_feat) == len(single_query_feat)
print('Write to mat file...')
import scipy.io as sio
if not os.path.exists('./matdata'):
os.mkdir('./matdata')
sio.savemat('./matdata/trainID.mat',
{'trainID': np.array(single_train_ids)})
sio.savemat('./matdata/queryID.mat',
{'queryID': np.array(single_query_ids)})
sio.savemat('./matdata/trainCAM.mat',
{'trainCAM': np.array(single_train_cams)})
sio.savemat('./matdata/queryCAM.mat',
{'queryCAM': np.array(single_query_cams)})
sio.savemat('./matdata/testID.mat', {'testID': np.array(gallery_ids)})
sio.savemat('./matdata/testCAM.mat', {'testCAM': np.array(gallery_cams)})
sio.savemat('./matdata/Hist_train.mat',
{'Hist_train': np.array(single_train_feat)})
sio.savemat('./matdata/Hist_query.mat',
{'Hist_query': np.array(single_query_feat)})
sio.savemat('./matdata/Hist_test.mat',
{'Hist_test': np.array(gallery_feat)})
sio.savemat('./matdata/Hist_query_max.mat',
{'Hist_max': np.array(query_max_feat)})
sio.savemat('./matdata/Hist_query_avg.mat',
{'Hist_avg': np.array(query_avg_feat)})
return
def reid_draw(frame, b_b, model, cfg, shujuku, threadPubMsg, camera_id):
global size
# print("size = ", size)
# print("b_b = ", b_b)
# print model
id_name = 0
cfg.cuda()
left = int((b_b[0] - b_b[2] / 2.0) * size[0])
top = int((b_b[1] - b_b[3] / 2.0) * size[1])
right = int((b_b[0] + b_b[2] / 2.0) * size[0])
bottom = int((b_b[1] + b_b[3] / 2.0) * size[1])
if top >= bottom or left >= right or top <= 0 or left <= 0:
return 0, 0, 0, 0, 0
ratio = float(bottom - top) / (right - left)
# print(ratio)
# if ratio < 2.0:
# # print('filtered out')
# return left, top, right, bottom, 999
if bottom > 530:
ratio = float(bottom - top) / (right - left)
#print("ratio is: {}".format(ratio))
if ratio < 1.5:
#print("ratio is: {}".format(ratio))
#print('filtered out')
return left, top, right, bottom, 999
img1 = jieduan(frame, left, top, right, bottom)
img = preprocess(img1)
feature = extract_cnn_feature(model, img.cuda())
minsim = -1
# id_name = 1
rentidir = '/home/tujh/renti/'
# for feature2,filename in shujuku:
for query in shujuku:
for fea in shujuku[query]:
distan = pairwise_distance(feature, fea)
if minsim > distan or minsim == -1:
minsim = distan
id_name = query
'''
distan=0
discount=0
for fea in shujuku[query]:
distan += pairwise_distance(feature,fea)
discount+=1
if discount!=0:
distan=distan/discount
if minsim > distan or minsim == -1 or distan!=0:
minsim = distan
id_name = query
'''
cv2.rectangle(frame, (left, top), (right, bottom), (255, 0, 0), 2)
cv2.putText(frame, str(id_name), (left, top), cv2.FONT_HERSHEY_COMPLEX, 6,
(255, 0, 0), 2)
# print(frame.shape)
# left_x = 450
# top_y = 30
# right_m = 570
# bottom_n = 130
left_x = 86
top_y = 130
right_m = 413
bottom_n = 479
#cv2.rectangle(frame, (left_x, top_y), (right_m, bottom_n), (0, 255, 0), 2)
calcIOU1 = calcIOU(left, top, right, bottom, left_x, top_y, right_m,
bottom_n)
print("calIOU : {}".format(calcIOU1))
if calcIOU1 > 0.8:
flag_check = 1
else:
flag_check = 0
print('flag', flag_check)
print(calcIOU1)
print("set customer message")
customer_name = "name" + str(id_name)
customer_id = id_name # number
global isAskPerson
print(isAskPerson)
# if isAskPerson == True and flag_check==1:
# cv2.imwrite('/home/zhaocy/catkin_touching_AI/checkout_img/checkout_' + str(customer_id) + '.jpg')
if flag_check == 1:
threadPubMsg.set_customer(name=customer_name,
person_id=customer_id,
camera_id=camera_id,
x=left,
y=top,
w=right,
h=bottom)
threadPubMsg.set_commodity_recognition_trigger(camera_id=camera_id,
person_id=customer_id,
flag=1,
flag1=0,
flag2=0)
print('save_images')
cv2.imwrite(
'/home/zhaocy/catkin_touching_AI/checkout_img/checkout_' +
str(customer_id) + '.jpg', frame)
else:
# print(1111111111111)
threadPubMsg.set_customer(name=customer_name,
person_id=customer_id,
camera_id=0,
x=left,
y=top,
w=right,
h=bottom)
threadPubMsg.set_commodity_recognition_trigger(camera_id=camera_id,
person_id=customer_id,
flag=0,
flag1=0,
flag2=0)
# return frame
return left, right, top, bottom, id_name
