def distance_test(self):
data_loader = LoadImages(
'/home/smile/cv/person_search_demo/data/test1',
img_size=416,
half=False)
imgs = []
paths = []
for index, (path, img, im0, vid_cap) in enumerate(data_loader):
img = Image.fromarray(cv2.cvtColor(
im0, cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
img = build_transforms(reidCfg)(img).unsqueeze(0) # torch.Size
imgs.append(img)
paths.append(path)
feats = self.reid_person(imgs)
query_feats = feats
gallery_feats = feats
m, n = query_feats.shape[0], gallery_feats.shape[0]
visual_distmat = torch.pow(query_feats, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t()
visual_distmat.addmm_(1, -2, query_feats, gallery_feats.t())
visual_distmat = visual_distmat.cpu()
res = visual_distmat.numpy().tolist()
res = paths.append(res)
print('\n')
print(res)
def load_img(cfg):
# 验证集的预处理
val_transforms = build_transforms(cfg)
num_workers = cfg.DATALOADER.NUM_WORKERS # 加载图像进程数 8
dataset = Market1501(root=cfg.DATASETS.ROOT_DIR)
print("dataset:", type(dataset), dataset)
print("dataset.query:", type(dataset.query), dataset.query)
val_set = ImageDataset(dataset.query, val_transforms)
val_loader = DataLoader(
val_set, batch_size=cfg.TEST.IMS_PER_BATCH, shuffle=False, num_workers=num_workers,
collate_fn=val_collate_fn
)
return val_loader, len(dataset.query)
def detect_person(self, path, img, im0, vid_cap):
imgs = []
locs = []
# Get detections shape: (3, 416, 320)
img = torch.from_numpy(img).unsqueeze(0).to(
self.device) # torch.Size([1, 3, 416, 320])
pred, _ = self.detection_model(img) # 经过处理的网络预测
det = non_max_suppression(pred.float(), self.conf_thres,
self.nms_thres)[0] # torch.Size([5, 7])
if det is not None and len(det) > 0:
# Rescale boxes from 416 to true image size 映射到原图
det[:, :4] = scale_coords(img.shape[2:], det[:, :4],
im0.shape).round()
print('%gx%g ' % img.shape[2:], end='')
# 对图片的所有类进行遍历循环，得到了当前类别的个数
for c in det[:, -1].unique():
n = (det[:, -1] == c).sum()
if self.classes[int(c)] == 'person':
print('%g %ss' % (n, self.classes[int(c)]), end=', ')
for *xyxy, conf, cls_conf, cls in det: # 对于最后的预测框进行遍历
# *xyxy: 对于原图来说的左上角右下角坐标: [tensor(349.), tensor(26.), tensor(468.), tensor(341.)]
if self.classes[int(cls)] == 'person':
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
# 如果检测到的行人太小了，感觉意义也不大
# 这里需要根据实际情况稍微设置下
if w * h > self.bbox_scale_thres:
locs.append((xmin, ymin, xmax, ymax))
crop_img = im0[ymin:ymax,
xmin:xmax] # HWC (602, 233, 3)
crop_img = Image.fromarray(
cv2.cvtColor(crop_img,
cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(
crop_img).unsqueeze(
0) # torch.Size([1, 3, 256, 128])
imgs.append(crop_img)
return pred, imgs, locs
def detect(
cfg,
data,
weights,
images='data/samples', # input folder
output='output', # output folder
fourcc='mp4v', # video codec
img_size=416,
conf_thres=0.25,
nms_thres=0.4,
dist_thres=1.3,
save_txt=False,
save_images=True):
# Initialize
device = torch_utils.select_device(force_cpu=False)
torch.backends.cudnn.benchmark = False # set False for reproducible results
if os.path.exists(output):
shutil.rmtree(output) # delete output folder
os.makedirs(output) # make new output folder
############# 行人重识别模型初始化 #############
query_loader, num_query = make_data_loader(reidCfg)
reidModel = build_model(reidCfg, num_classes=10126)
reidModel.load_param(reidCfg.TEST.WEIGHT)
reidModel.to(device).eval()
print('num query %d' % num_query)
query_feats = defaultdict(list)
query_pids = []
for i, batch in enumerate(query_loader):
with torch.no_grad():
img, pid, camid = batch
img = img.to(device)
feat = reidModel(
img) # 一共2张待查询图片，每张图片特征向量2048 torch.Size([2, 2048])
for j, f in enumerate(feat):
if (not pid[j] in query_pids):
query_pids.append(pid[j])
print(f.cpu().numpy())
query_feats[pid[j]].append(f.cpu().numpy())
for pid in query_pids:
temp = np.array(query_feats[pid])
print(temp)
query_feats[pid] = torch.from_numpy(temp).float().to(
device) # torch.Size([2, 2048])
print(query_feats[pid])
query_feats[pid] = torch.nn.functional.normalize(query_feats[pid],
dim=1,
p=2) # 计算出查询图片的特征向量
print(query_feats[pid])
print("The query feature is normalized")
############# 行人检测模型初始化 #############
model = Darknet(cfg, img_size)
# Load weights
if weights.endswith('.pt'): # pytorch format
model.load_state_dict(
torch.load(weights, map_location=device)['model'])
else: # darknet format
_ = load_darknet_weights(model, weights)
# Eval mode
model.to(device).eval()
# Half precision
opt.half = opt.half and device.type != 'cpu' # half precision only supported on CUDA
if opt.half:
model.half()
# Set Dataloader
vid_path, vid_writer = None, None
if opt.webcam:
save_images = False
dataloader = LoadWebcam(img_size=img_size, half=opt.half)
else:
dataloader = LoadImages(images, img_size=img_size, half=opt.half)
# Get classes and colors
# parse_data_cfg(data)['names']:得到类别名称文件路径 names=data/coco.names
classes = load_classes(
parse_data_cfg(data)['names']) # 得到类别名列表: ['person', 'bicycle'...]
colors = [[random.randint(0, 255) for _ in range(3)]
for _ in range(len(classes))] # 对于每种类别随机使用一种颜色画框
# Run inference
t0 = time.time()
for i, (path, img, im0, vid_cap) in enumerate(dataloader):
t = time.time()
# if i < 500 or i % 5 == 0:
# continue
save_path = str(Path(output) / Path(path).name) # 保存的路径
# Get detections shape: (3, 416, 320)
img = torch.from_numpy(img).unsqueeze(0).to(
device) # torch.Size([1, 3, 416, 320])
pred, _ = model(img) # 经过处理的网络预测，和原始的
det = non_max_suppression(pred.float(), conf_thres,
nms_thres)[0] # torch.Size([5, 7])
if det is not None and len(det) > 0:
# Rescale boxes from 416 to true image size 映射到原图
det[:, :4] = scale_coords(img.shape[2:], det[:, :4],
im0.shape).round()
# Print results to screen image 1/3 data\samples\000493.jpg: 288x416 5 persons, Done. (0.869s)
print('%gx%g ' % img.shape[2:],
end='') # print image size '288x416'
for c in det[:, -1].unique(): # 对图片的所有类进行遍历循环
n = (det[:, -1] == c).sum() # 得到了当前类别的个数，也可以用来统计数目
if classes[int(c)] == 'person':
print('%g %ss' % (n, classes[int(c)]),
end=', ') # 打印个数和类别'5 persons'
# Draw bounding boxes and labels of detections
# (x1y1x2y2, obj_conf, class_conf, class_pred)
count = 0
gallery_img = []
gallery_loc = []
for *xyxy, conf, cls_conf, cls in det: # 对于最后的预测框进行遍历
# *xyxy: 对于原图来说的左上角右下角坐标: [tensor(349.), tensor(26.), tensor(468.), tensor(341.)]
if save_txt: # Write to file
with open(save_path + '.txt', 'a') as file:
file.write(('%g ' * 6 + '\n') % (*xyxy, cls, conf))
# Add bbox to the image
label = '%s %.2f' % (classes[int(cls)], conf) # 'person 1.00'
if classes[int(cls)] == 'person':
#plot_one_bo x(xyxy, im0, label=label, color=colors[int(cls)])
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
# 如果检测到的行人太小了，感觉意义也不大
# 这里需要根据实际情况稍微设置下
if w * h > 500:
gallery_loc.append((xmin, ymin, xmax, ymax))
crop_img = im0[ymin:ymax,
xmin:xmax] # HWC (602, 233, 3)
crop_img = Image.fromarray(
cv2.cvtColor(crop_img,
cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(
crop_img).unsqueeze(
0) # torch.Size([1, 3, 256, 128])
gallery_img.append(crop_img)
if gallery_img:
gallery_img = torch.cat(gallery_img,
dim=0) # torch.Size([7, 3, 256, 128])
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(gallery_img) # torch.Size([7, 2048])
print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(
gallery_feats, dim=1, p=2) # 计算出查询图片的特征向量
for pid in query_pids:
m, n = query_feats[pid].shape[0], gallery_feats.shape[0]
distmat = torch.pow(query_feats[pid], 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, query_feats[pid], gallery_feats.t())
# distmat = (qf - gf)^2
# distmat = np.array([[1.79536, 2.00926, 0.52790, 1.98851, 2.15138, 1.75929, 1.99410],
# [1.78843, 1.96036, 0.53674, 1.98929, 1.99490, 1.84878, 1.98575]])
distmat = distmat.cpu().numpy() # <class 'tuple'>: (3, 12)
distmat = distmat.sum(axis=0) / len(
query_feats[pid]) # 平均一下query中同一行人的多个结果
index = distmat.argmin()
if distmat[index] < dist_thres:
print('距离：%s' % distmat[index])
plot_one_box(gallery_loc[index],
im0,
label=str(pid),
color=colors[int(cls)])
# cv2.imshow('person search', im0)
# cv2.waitKey()
print('Done. (%.3fs)' % (time.time() - t))
if opt.webcam: # Show live webcam
cv2.imshow(weights, im0)
if save_images: # Save image with detections
if dataloader.mode == 'images':
cv2.imwrite(save_path, im0)
else:
if vid_path != save_path: # new video
vid_path = save_path
if isinstance(vid_writer, cv2.VideoWriter):
vid_writer.release() # release previous video writer
fps = vid_cap.get(cv2.CAP_PROP_FPS)
width = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
vid_writer = cv2.VideoWriter(
save_path, cv2.VideoWriter_fourcc(*fourcc), fps,
(width, height))
vid_writer.write(im0)
if save_images:
print('Results saved to %s' % os.getcwd() + os.sep + output)
if platform == 'darwin': # macos
os.system('open ' + output + ' ' + save_path)
print('Done. (%.3fs)' % (time.time() - t0))
def detect(
cfg,
data,
weights,
images='data/samples', # input folder
output='output', # output folder
fourcc='mp4v', # video codec
img_size=416,
conf_thres=0.5,
nms_thres=0.5,
dist_thres=1.0,
save_txt=False,
save_images=True):
# Initialize
device = torch_utils.select_device(force_cpu=False) # 不强制使用cpu
print("using device:", type(device), device)
torch.backends.cudnn.benchmark = False # set False for reproducible results
if os.path.exists(output):
shutil.rmtree(output) # delete output folder
os.makedirs(output) # make new output folder
############# 加载<pid, 中文名, 颜色>列表 #############
pidNameInfo = make_pidNames_loader(reidCfg)
print("pidNameInfo:", pidNameInfo)
############# 行人重识别模型初始化 #############
query_loader, num_query = make_data_loader(reidCfg)
print("query_loader:", type(query_loader), query_loader)
reidModel = build_model(reidCfg, num_classes=10126)
reidModel.load_param(reidCfg.TEST.WEIGHT)
reidModel.to(device).eval()
query_feats = []
query_pids = []
for i, batch in enumerate(query_loader):
with torch.no_grad():
img, pid, camid = batch # 图像，行人ID，相机ID
img = img.to(device)
feat = reidModel(
img) # 一共2张待查询图片，每张图片特征向量2048 torch.Size([2, 2048])
print("feat:", type(feat), feat.shape, feat)
query_feats.append(feat)
query_pids.extend(np.asarray(
pid)) # extend() 函数用于在列表末尾一次性追加另一个序列中的多个值（用新列表扩展原来的列表）。
query_feats = torch.cat(query_feats, dim=0) # torch.Size([2, 2048])
print("The query feature is normalized")
query_feats = torch.nn.functional.normalize(query_feats, dim=1,
p=2) # 计算出查询图片的特征向量
############# 行人检测模型初始化 #############
model = Darknet(cfg, img_size)
# Load weights
if weights.endswith('.pt'): # pytorch format
model.load_state_dict(
torch.load(weights, map_location=device)['model'])
# model.load_state_dict(torch.load(weights, map_location='cpu')['model']) # 使用cpu
else: # darknet format
_ = load_darknet_weights(model, weights)
# Eval mode
model.to(device).eval()
# Half precision
opt.half = opt.half and device.type != 'cpu' # half precision only supported on CUDA
if opt.half:
model.half()
# Set Dataloader
vid_path, vid_writer = None, None
if opt.webcam:
save_images = False
dataloader = LoadWebcam(img_size=img_size, half=opt.half)
else:
dataloader = LoadImages(images, img_size=img_size, half=opt.half)
# Get classes and colors
# parse_data_cfg(data)['names']:得到类别名称文件路径 names=data/coco.names
classes = load_classes(
parse_data_cfg(data)['names']) # 得到类别名列表: ['person', 'bicycle'...]
# colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(classes))] # 对于每种类别随机使用一种颜色画框
# colors = [[40, 92, 230] for _ in range(len(classes))] # 只检测人，使用相同的颜色
# Run inference
t0 = time.time()
for i, (path, img, im0, vid_cap) in enumerate(dataloader):
t = time.time()
# if i < 500 or i % 5 == 0:
# continue
save_path = str(Path(output) / Path(path).name) # 保存的路径
# Get detections shape: (3, 416, 320)
img = torch.from_numpy(img).unsqueeze(0).to(
device) # torch.Size([1, 3, 416, 320])
pred, _ = model(img) # 经过处理的网络预测，和原始的
det = non_max_suppression(pred.float(), conf_thres,
nms_thres)[0] # torch.Size([5, 7])
if det is not None and len(det) > 0:
# Rescale boxes from 416 to true image size 映射到原图
det[:, :4] = scale_coords(img.shape[2:], det[:, :4],
im0.shape).round()
# Print results to screen image 1/3 data\samples\000493.jpg: 288x416 5 persons, Done. (0.869s)
print('%gx%g ' % img.shape[2:],
end='') # print image size '288x416'
for c in det[:, -1].unique(): # 对图片的所有类进行遍历循环
n = (det[:, -1] == c).sum() # 得到了当前类别的个数，也可以用来统计数目
if classes[int(c)] == 'person':
print('%g %ss' % (n, classes[int(c)]),
end=', ') # 打印个数和类别'5 persons'
# Draw bounding boxes and labels of detections
# (x1y1x2y2, obj_conf, class_conf, class_pred)
count = 0
gallery_img = []
gallery_loc = []
for *xyxy, conf, cls_conf, cls in det: # 对于最后的预测框进行遍历
# *xyxy: 对于原图来说的左上角右下角坐标: [tensor(349.), tensor(26.), tensor(468.), tensor(341.)]
if save_txt: # Write to file
with open(save_path + '.txt', 'a') as file:
file.write(('%g ' * 6 + '\n') % (*xyxy, cls, conf))
# Add bbox to the image
label = '%s %.2f' % (classes[int(cls)], conf) # 'person 1.00'
if classes[int(cls)] == 'person':
#plot_one_bo x(xyxy, im0, label=label, color=colors[int(cls)])
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
# 如果检测到的行人太小了，感觉意义也不大
# 这里需要根据实际情况稍微设置下
if w * h > 500:
gallery_loc.append((xmin, ymin, xmax, ymax))
crop_img = im0[ymin:ymax,
xmin:xmax] # HWC (602, 233, 3)
crop_img = Image.fromarray(
cv2.cvtColor(crop_img,
cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(
crop_img).unsqueeze(
0) # torch.Size([1, 3, 256, 128])
gallery_img.append(crop_img)
if gallery_img:
gallery_img = torch.cat(gallery_img,
dim=0) # torch.Size([7, 3, 256, 128])
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(gallery_img) # torch.Size([7, 2048])
print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(
gallery_feats, dim=1, p=2) # 计算出查询图片的特征向量
# m: 2，待查询人的个数
# n: 7，yolo检测到人的个数
m, n = query_feats.shape[0], gallery_feats.shape[0]
# print("query_feats.shape:", query_feats.shape, "gallery_feats.shape:", gallery_feats.shape)
distmat = torch.pow(query_feats, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t() # .t()，矩阵转置
# out=(beta∗M)+(alpha∗mat1@mat2)
# qf^2 + gf^2 - 2 * [email protected]()
# distmat - 2 * [email protected]()
# distmat: qf^2 + gf^2
# qf: torch.Size([2, 2048])
# gf: torch.Size([7, 2048])
distmat.addmm_(1, -2, query_feats,
gallery_feats.t()) # distmat行数：待检人数；列数：库图片中人数
# distmat = (qf - gf)^2
# distmat = np.array([[1.79536, 2.00926, 0.52790, 1.98851, 2.15138, 1.75929, 1.99410],
# [1.78843, 1.96036, 0.53674, 1.98929, 1.99490, 1.84878, 1.98575]])
distmat = distmat.cpu().numpy(
) # <class 'tuple'>: (3, 12)，cuda转到内存中
print("distmat:", type(distmat), m, n, distmat.shape)
pidMatDict = splitDistmat(query_pids,
distmat) # 按pid拆分，按行拆分，<pid, mat>
pid_freq = Counter(query_pids)
for k in pidMatDict.keys(): # k表示pid
tmp_mat = pidMatDict[k] # 某一个人的矩阵
tmp_mat = tmp_mat.sum(
axis=0) / pid_freq[k] # 平均一下query中同一行人的多个结果
# tmp_mat = tmp_mat.sum(axis=0)
index = tmp_mat.argmin() # 返回距离最小值的下标（图中哪个人最像待检测的人）
# print("tmp_mat:", type(tmp_mat), m, n, tmp_mat.shape, tmp_mat)
print('距 离：%s' % tmp_mat[index], index, dist_thres, n)
# print("gallery_loc:", gallery_loc)
if tmp_mat[index] < dist_thres:
print('距离：%s' % tmp_mat[index])
# # cv2显示中文出现乱码，需转为PIL添加中文后再转回来。这样在方法中调用的话最后显示为颜色涂层filled，没有内容。
# # 解决办法：把画框和添加中文label过程直接放到这里来，避免函数调用式转写。（英文可以直接调用该方法，不会出现乱码）
# plot_one_box(gallery_loc[index], im0, label='%s:%s' % (pidNameInfo[str(k)][1], tmp_mat[index]), color=getColorArr(pidNameInfo[str(k)][2]))
# 准备plot_one_box()方法的各项参数
x = gallery_loc[index] # 画框范围（左上、右下坐标）
label = '%s:%s' % (pidNameInfo[str(k)][1],
tmp_mat[index]) # 要标记的内容（中文）
# label = '%s' % pidNameInfo[str(k)][0]
color = getColorArr(pidNameInfo[str(k)][2]) # 画框颜色
tl = round(0.002 * (im0.shape[0] + im0.shape[1]) /
2) + 1 # line thickness，线条厚度
color = color or [
random.randint(0, 255) for _ in range(3)
] # 如果没指定颜色，则随机
c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])
) # 左上xy，右下xy
cv2.rectangle(im0, c1, c2, color, thickness=tl)
if label:
# 对于内容框，c1为左下，c2为右上
tf = max(tl - 1, 1) # font thickness，字体粗细
# t_size = cv2.getTextSize(label, fontFace=0, fontScale=tl / 3, thickness=tf)[0] # fontFace=1时，((310, 22), 10)
# t_size = cv2.getTextSize(label, fontFace=1, fontScale=tl / 3, thickness=tf)[0] # fontFace=1时，((156, 10), 6)
zh_cn_nums = get_zhcn_number(
label) # 中文的字数（一个中文字20个像素宽，一个英文字10个像素宽）
t_size = (20 * zh_cn_nums + 10 *
(len(label) - zh_cn_nums), 22)
c2 = c1[0] + t_size[0], c1[1] - t_size[
1] - 3 # 纵坐标，多减3目的是字上方稍留空
cv2.rectangle(im0, c1, c2, color, -1) # filled
print("t_size:", t_size, " c1:", c1, " c2:", c2)
# Draw a label with a name below the face
# cv2.rectangle(im0, c1, c2, (0, 0, 255), cv2.FILLED)
font = cv2.FONT_HERSHEY_DUPLEX
# 将CV2转为PIL，添加中文label后再转回来
pil_img = Image.fromarray(
cv2.cvtColor(im0, cv2.COLOR_BGR2RGB))
draw = ImageDraw.Draw(pil_img)
font = ImageFont.truetype('simhei.ttf',
20,
encoding='utf-8')
draw.text((c1[0], c1[1] - 20),
label, (255, 255, 255),
font=font)
im0 = cv2.cvtColor(np.array(pil_img),
cv2.COLOR_RGB2BGR) # PIL转CV2
# cv2.imshow('person search', im0)
# cv2.waitKey()
print('Done. (%.3fs)' % (time.time() - t))
if opt.webcam: # Show live webcam
cv2.imshow(weights, im0)
if save_images: # Save image with detections
if dataloader.mode == 'images':
cv2.imwrite(save_path, im0)
else:
if vid_path != save_path: # new video
vid_path = save_path
if isinstance(vid_writer, cv2.VideoWriter):
vid_writer.release() # release previous video writer
fps = vid_cap.get(cv2.CAP_PROP_FPS)
width = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
vid_writer = cv2.VideoWriter(
save_path, cv2.VideoWriter_fourcc(*fourcc), fps,
(width, height))
vid_writer.write(im0)
if save_images:
print('Results saved to %s' % os.getcwd() + os.sep + output)
if platform == 'darwin': # macos
os.system('open ' + output + ' ' + save_path)
print('Done. (%.3fs)' % (time.time() - t0))
def yolo_frames(unique_name):
cam_id = unique_name[1]
device = torch_utils.select_device(force_cpu=False)
torch.backends.cudnn.benchmark = False # set False for reproducible results
cfg = 'cfg/yolov3.cfg'
data = 'data/coco.data'
weights = 'weights/yolov3.weights'
half = False
img_size = 416
conf_thres = 0.25
nms_thres = 0.4
dist_thres = 1.4
#reid 모델 생성
query_loader, num_query = make_data_loader(reidCfg)
reidModel = build_model(reidCfg, num_classes=10126)
reidModel.load_param(reidCfg.TEST.WEIGHT)
reidModel.to(device).eval()
#reid하기 위한 query 정보
query_feats = defaultdict(list)
query_pids = []
#query 정보 가져오기
for i, batch in enumerate(query_loader):
with torch.no_grad():
img, pid, camid = batch
img = img.to(device)
feat = reidModel(img)
for j, f in enumerate(feat):
if (not pid[j] in query_pids):
query_pids.append(pid[j])
query_feats[pid[j]].append(f.cpu().numpy())
#query 정보 torch형식으로 변경
for pid in query_pids:
temp = np.array(query_feats[pid])
query_feats[pid] = torch.from_numpy(temp).float().to(device)
query_feats[pid] = torch.nn.functional.normalize(query_feats[pid],
dim=1,
p=2)
print("The query feature is normalized")
model = Darknet(cfg, img_size) #config로 디텍션 모델 생성
# Load weights
if weights.endswith('.pt'): # pytorch format
model.load_state_dict(
torch.load(weights, map_location=device)['model'])
else: # darknet format
_ = load_darknet_weights(model, weights)
# Eval mode
model.to(device).eval()
# Half precision
half = half and device.type != 'cpu' # half precision only supported on CUDA
if half:
model.half()
# Set Dataloader
dataloader = LoadWebcam(cam_id, img_size=img_size, half=half)
# Get classes and colors
# parse_data_cfg(data)['names'] names=data/coco.names
classes = load_classes(parse_data_cfg(
data)['names']) # 코코 네임 파일에서 클래스 전부 가져옴 ['person', 'bicycle'...]
colors = [[random.randint(0, 255) for _ in range(3)]
for _ in range(len(classes))] # 이 클래스 수대로 박스의 색을 random으로 생성
#count = 0 #이미지 자르기 위한 카운트
# Run inference
t0 = time.time()
for i, (path, img, im0, vid_cap) in enumerate(dataloader):
patient_map = map.Map
now = time.localtime()
# If the saved minute is not same with the current minute, it means this camera is the first one that access to the map information in this minute.
# Since we have to clear the second array every minute, change the minute and clear the second array.
if (now.tm_min != patient_map.minute):
patient_map.minute = now.tm_min
patient_map.sec_array.clear()
# If there is no information about current second, it means this is the first access to the map in this second.
# We should init the map information each second.
if (now.tm_sec not in patient_map.sec_array):
patient_map.sec_array.append(now.tm_sec)
patient_map.exist_id = []
patient_map.camera_map = {0: [], 1: [], 2: []}
patient_map.total_count = {0: 0, 1: 0, 2: 0}
if i % 5 != 0: #이미지 처리 부하 줄이기
continue
# Get detections shape: (3, 416, 320)
img = torch.from_numpy(img).unsqueeze(0).to(
device) # torch.Size([1, 3, 416, 320]) #이미지 torch 형식으로 바꾸기
pred, _ = model(img) #이미지 디텍션
det = non_max_suppression(pred.float(), conf_thres, nms_thres)[
0] # torch.Size([5, 7]) #threshold로 이미지 거른 후 det변수에 초기화
if det is not None and len(det) > 0:
# Rescale boxes from 416 to true image size
det[:, :4] = scale_coords(img.shape[2:], det[:, :4],
im0.shape).round()
# Print results to screen image 1/3 data\samples\000493.jpg: 288x416 5 persons, Done. (0.869s)
#print('%gx%g ' % img.shape[2:], end='') # print image size '288x416'
#for c in det[:, -1].unique(): # 对图片的所有类进行遍历循环
# n = (det[:, -1] == c).sum() # 得到了当前类别的个数，也可以用来统计数目
#if classes[int(c)] == 'person':
# print('%g %ss' % (n, classes[int(c)]), end=', ') # 打印个数和类别'5 persons'
# print(" ")
# Draw bounding boxes and labels of detections
# (x1y1x2y2, obj_conf, class_conf, class_pred)
gallery_img = [] #사람의 이미지만 따로 저장하는 Array
gallery_loc = [] #사진의 좌표를 저장하는 Array
for *xyxy, conf, cls_conf, cls in det: # det의 정보엔 박스의 좌표, 정확도, 클래스인지 확인하는 정확도가 있음
# *xyxy: [tensor(349.), tensor(26.), tensor(468.), tensor(341.)] tensor로 좌표가 있음
# Add bbox to the image
# label = '%s %.2f' % (classes[int(cls)], conf) # 'person 1.00'
if classes[int(
cls)] == 'person': # detect한 클래스가 person class라면
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
if w * h > 500:
gallery_loc.append((xmin, ymin, xmax, ymax))
crop_img = im0[ymin:ymax,
xmin:xmax] # HWC (602, 233, 3)
#cv2.imwrite('./temp/'+str(count)+'.jpg',crop_img) #query 이미지로 쓰기위해 temp 폴더에 내 이미지 저장
#count=count+1
crop_img = Image.fromarray(
cv2.cvtColor(
crop_img,
cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(
crop_img).unsqueeze(
0) # torch.Size([1, 3, 256, 128])
gallery_img.append(crop_img)
plot_one_box(xyxy, im0,
color=[128, 128,
128]) #사람을 흰색으로 박스 치기
map.Map.total_count[int(cam_id)] = len(gallery_img)
if gallery_img: #사람의 이미지만 자른 Array의 데이터가 존재하면
gallery_img = torch.cat(
gallery_img, dim=0) # torch.Size([7, 3, 256, 128])
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(
gallery_img) # torch.Size([7, 2048])
#print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(
gallery_feats, dim=1, p=2) # 计算出查询图片的特征向量
#그 이미지의 특징을 뽑아옴
# m: 2
# n: 7
for pid in query_pids:
m, n = query_feats[pid].shape[0], gallery_feats.shape[
0]
distmat = torch.pow(query_feats[pid], 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t()
# out=(beta∗M)+(alpha∗mat1@mat2)
# qf^2 + gf^2 - 2 * [email protected]()
# distmat - 2 * [email protected]()
# distmat: qf^2 + gf^2
# qf: torch.Size([2, 2048])
# gf: torch.Size([7, 2048])
distmat.addmm_(1, -2, query_feats[pid],
gallery_feats.t())
# distmat = (qf - gf)^2
# distma2 번 목표 찾음 2번 카메라：1.2738347
#t = np.array([[1.79536, 2.00926, 0.52790, 1.98851, 2.15138, 1.75929, 1.99410],
# [1.78843, 1.96036, 0.53674, 1.98929, 1.99490, 1.84878, 1.98575]])
distmat = distmat.cpu().detach().numpy(
) # <class 'tuple'>: (3, 12)
distmat = distmat.sum(axis=0) / len(
query_feats[pid]) # 쿼리의 특징과 현재 이미지의 특징의 차이를 계산
index = distmat.argmin()
if distmat[
index] < dist_thres: #그 차이가 위에서 지정한 treshold보다 작으면 일치하다
print('%i 번 목표 찾음 %s번 카메라：%s' %
(pid, cam_id, distmat[index]))
plot_one_box(gallery_loc[index],
im0,
label=str(pid),
color=[0, 0, 255])
if (pid not in patient_map.exist_id):
patient_map.exist_id.append(pid)
if (pid not in patient_map.camera_map[int(cam_id)]
):
patient_map.camera_map[int(cam_id)].append(pid)
#print("exist id : ", patient_map.exist_id)
#print(patient_map.camera_map)
print("total : ", len(gallery_img))
filename = time.strftime(
"%Y%m%d", time.localtime(time.time())
) + '_person' + str(pid) + '.txt'
f = open(filename, 'a')
f.write('\n' + cam_id + ' - ' +
time.strftime('%H : %M : %S'))
f.close
#If the map of this camera ID is still false, it means there was no identified query in this second.
#if(patient_map.camera_map[int(cam_id)] == False):
yield cam_id, im0
def search_detect(dataloader_item, model, reidModel, device, classes, colors,
weights):
global query_time_last
global query_time_now
global query_feats
conf_thres = 0.1
nms_thres = 0.4
dist_thres = 3.0
output = 'output'
fourcc = 'mp4v'
t = time.time()
path, img, im0, vid_cap = dataloader_item
# print(path, img.shape, im0.shape, vid_cap)
# print(aaa)
# data/samples/c1s1_001051.jpg (3, 320, 416) (480, 640, 3) None
vid_path, vid_writer = None, None
############# query初始化 #############
if len(os.listdir('query')) < 1:
print('not enough query')
return
else:
if query_time_now == query_time_last or query_time_now - query_time_last >= 1:
query_loader, num_query = make_data_loader(reidCfg)
query_feats = []
for i, batch in enumerate(query_loader):
with torch.no_grad():
img_q, pid, camid = batch
img_q = img_q.to(device)
feat = reidModel(
img_q) # 一共2张待查询图片，每张图片特征向量2048 torch.Size([2, 2048])
query_feats.append(feat)
query_feats = torch.cat(query_feats,
dim=0) # torch.Size([2, 2048])
query_feats = torch.nn.functional.normalize(query_feats,
dim=1,
p=2) # 计算出查询图片的特征向量
print("The query feature is normalized")
query_time_last = query_time_now
elif len(query_feats) == 0:
print('no query_feats')
return
query_time_now = time.time()
############# query初始化 END #############
if not os.path.exists(output):
os.makedirs(output)
save_path = str(Path(output) / Path(path).name) # 保存的路径
# Get detections shape: (3, 416, 320)
img = torch.from_numpy(img).unsqueeze(0).to(
device) # torch.Size([1, 3, 416, 320])
pred, _ = model(img) # 经过处理的网络预测，和原始的
det = non_max_suppression(pred.float(), conf_thres,
nms_thres)[0] # torch.Size([5, 7])
if det is not None and len(det) > 0:
# Rescale boxes from 416 to true image size 映射到原图
det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
# Print results to screen image 1/3 data\samples\000493.jpg: 288x416 5 persons, Done. (0.869s)
print('%gx%g ' % img.shape[2:], end='') # print image size '288x416'
for c in det[:, -1].unique(): # 对图片的所有类进行遍历循环
n = (det[:, -1] == c).sum() # 得到了当前类别的个数，也可以用来统计数目
if classes[int(c)] == 'person':
print('%g %ss' % (n, classes[int(c)]),
end=', ') # 打印个数和类别'5 persons'
# Draw bounding boxes and labels of detections
# (x1y1x2y2, obj_conf, class_conf, class_pred)
count = 0
gallery_img = []
gallery_loc = []
for *xyxy, conf, cls_conf, cls in det: # 对于最后的预测框进行遍历
# *xyxy: 对于原图来说的左上角右下角坐标: [tensor(349.), tensor(26.), tensor(468.), tensor(341.)]
'''Write to file'''
# with open(save_path + '.txt', 'a') as file:
# file.write(('%g ' * 6 + '\n') % (*xyxy, cls, conf))
# Add bbox to the image
label = '%s %.2f' % (classes[int(cls)], conf) # 'person 1.00'
if classes[int(cls)] == 'person':
#plot_one_bo x(xyxy, im0, label=label, color=colors[int(cls)])
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
# 如果检测到的行人太小了，感觉意义也不大
# 这里需要根据实际情况稍微设置下
# if h>2*w and h*w > 100*50:
if h > 100 and w > 50:
gallery_loc.append((xmin, ymin, xmax, ymax))
crop_img = im0[ymin:ymax, xmin:xmax] # HWC (602, 233, 3)
crop_img = Image.fromarray(
cv2.cvtColor(crop_img,
cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(crop_img).unsqueeze(
0) # torch.Size([1, 3, 256, 128])
gallery_img.append(crop_img)
'''flip image and box'''
im0 = cv2.flip(im0, 1)
if gallery_img:
gallery_img = torch.cat(gallery_img,
dim=0) # torch.Size([7, 3, 256, 128])
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(gallery_img) # torch.Size([7, 2048])
gallery_feats = torch.nn.functional.normalize(gallery_feats,
dim=1,
p=2) # 计算出查询图片的特征向量
print("The gallery feature is normalized")
# m: 2
# n: 7
m, n = query_feats.shape[0], gallery_feats.shape[0]
distmat = torch.pow(query_feats, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t()
# out=(beta∗M)+(alpha∗mat1@mat2)
# qf^2 + gf^2 - 2 * [email protected]()
# distmat - 2 * [email protected]()
# distmat: qf^2 + gf^2
# qf: torch.Size([2, 2048])
# gf: torch.Size([7, 2048])
distmat.addmm_(1, -2, query_feats, gallery_feats.t())
# distmat = (qf - gf)^2
# distmat = np.array([[1.79536, 2.00926, 0.52790, 1.98851, 2.15138, 1.75929, 1.99410],
# [1.78843, 1.96036, 0.53674, 1.98929, 1.99490, 1.84878, 1.98575]])
distmat = distmat.cpu().numpy() # <class 'tuple'>: (3, 12)
distmat = distmat.sum(axis=0) / len(
query_feats) # 平均一下query中同一行人的多个结果
index = distmat.argmin()
if distmat[index] < dist_thres:
print('距离：%s' % distmat[index])
# print(gallery_loc[index])
xmin = im0.shape[1] - gallery_loc[index][2]
ymin = im0.shape[0] - gallery_loc[index][3]
xmax = im0.shape[1] - gallery_loc[index][0]
ymax = im0.shape[0] - gallery_loc[index][1]
# plot_one_box(gallery_loc[index], im0, label='find!', color=colors[int(cls)])
plot_one_box((xmin, ymin, xmax, ymax),
im0,
label='find!',
color=colors[int(cls)])
# cv2.imshow('person search', im0)
# cv2.waitKey()
print('Done. (%.3fs)' % (time.time() - t))
'''show image'''
# cv2.imshow(weights, im0)
'''save image'''
# cv2.imwrite(save_path, im0)
'''save webcam'''
# if vid_path != save_path: # new video
# vid_path = save_path
# if isinstance(vid_writer, cv2.VideoWriter):
# vid_writer.release() # release previous video writer
# fps = vid_cap.get(cv2.CAP_PROP_FPS)
# width = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
# height = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
# vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*fourcc), fps, (width, height))
# vid_writer.write(im0)
return im0
def detect(self, images, stamp, dist_thres=1.0):
camid = 0 #这里定义camid
# start re-id
crop_img = build_transforms(reidCfg)(images).unsqueeze(0)
gallery_img = []
gallery_img.append(crop_img)
if gallery_img:
gallery_img = torch.cat(gallery_img, dim=0)
gallery_img = gallery_img.to(self.device)
gallery_feats = self.reidModel(gallery_img)
print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(gallery_feats,
dim=1,
p=2)
for i, m_query_feat in enumerate(self.query_feats):
m_query_feat = torch.reshape(m_query_feat, (1, 2048))
m, n = m_query_feat.shape[0], gallery_feats.shape[0]
distmat = torch.pow(m_query_feat, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, m_query_feat, gallery_feats.t())
distmat = distmat.cpu().numpy()
distmat = distmat.sum(axis=0) / len(m_query_feat)
index = distmat.argmin()
print('stamp:', stamp, '\tpid:', self.query_pids[i],
'\tdismat:', distmat[index])
if distmat[index] < dist_thres:
return self.query_pids[i]
# 如果query中没有要找的目标
#Step 1:找到最大值，存储图片
if len(self.query_pids) != 0:
pids_copy = self.query_pids[:] #copy一下pids列表用来寻找最大值，以便分配新id
pids_copy.sort()
x = pids_copy[-1] + 1
else:
x = 0
x_str = str(x)
path = './query/' + x_str.rjust(4, '0') + '_c1s1_0_0_' + str(
time.time()) + '.jpg'
images.save(path)
#Step 2:自学习
with torch.no_grad():
#提取特征并normalized
new_id_img = build_transforms(reidCfg)(images).unsqueeze(0)
new_id_imgs = []
new_id_imgs.append(new_id_img)
new_id_imgs = torch.cat(new_id_imgs, dim=0)
new_id_imgs = new_id_imgs.to(self.device)
new_id_feats = self.reidModel(new_id_imgs)
new_id_feats = torch.nn.functional.normalize(new_id_feats,
dim=1,
p=2)
print("The gallery feature is normalized")
#加入总的特征矩阵
self.query_feats = torch.cat((self.query_feats, new_id_feats),
0)
temp = []
temp.append(x)
self.query_pids.extend(np.asarray(temp))
return x
from utils.utils import *
from reid.data import make_data_loader
from reid.data.transforms import build_transforms
from reid.modeling import build_model
from reid.config import cfg as reidCfg
import numpy as np
device = torch_utils.select_device(force_cpu=False)
torch.backends.cudnn.benchmark = False
reidCfg.DATASETS.ROOT_DIR = ('./query/')
query_loader, num_query = make_data_loader(reidCfg)
reidModel = build_model(reidCfg, num_classes=10126)
reidModel.load_param(reidCfg.TEST.WEIGHT)
reidModel.to(device).eval()
img = Image.open('./query/0001_c1s1_0_146.jpg')
crop_img = build_transforms(reidCfg)(img).unsqueeze(0)
gallery_img = []
gallery_img.append(crop_img)
gallery_img = torch.cat(gallery_img, dim=0)
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(gallery_img)
print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(gallery_feats, dim=1, p=2)
print(gallery_feats)
#plot_one_bo x(xyxy, im0, label=label, color=colors[int(cls)])
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
# 如果检测到的行人太小了，感觉意义也不大
# 这里需要根据实际情况稍微设置下
if h > 2 * w and h * w > 300 * 150:
print(h, w)
crop_img = im0[ymin:ymax, xmin:xmax] # HWC (602, 233, 3)
crop_img = Image.fromarray(
cv2.cvtColor(crop_img,
cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(crop_img).unsqueeze(
0) # torch.Size([1, 3, 256, 128])
gallery_img.append(crop_img)
if gallery_img:
gallery_img = torch.cat(
gallery_img, dim=0) # torch.Size([7, 3, 256, 128])
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(
gallery_img) # torch.Size([7, 2048])
print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(
gallery_feats, dim=1, p=2) # 计算出查询图片的特征向量
# m: 2
# n: 7
m, n = query_feats.shape[0], gallery_feats.shape[0]