def obj_detect(image):
input_size = cfg['input_size']
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
original_image = image
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image), [input_size, input_size])
image_data = image_data[np.newaxis, ...]
# model infer
pred_sbbox, pred_mbbox, pred_lbbox = obj_detect_infer.infer(image_data)
# post process get final bboxes
num_classes = cfg['num_classes']
pred_bbox = np.concatenate([np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))], axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size, input_size, 0.5)
bboxes = utils.nms(bboxes, 0.45, method='nms')
res = {}
objs = []
for bbox in bboxes:
objs.append({'x1':bbox[0], 'y1':bbox[1], 'x2':bbox[2], 'y2':bbox[3], 'confidence':bbox[4], 'class':bbox[5]})
res['objs'] = objs
return res
def predict(self, image):
org_image = np.copy(image)
org_h, org_w, _ = org_image.shape
# 将 image 转成需要的格式和 大小
image_data = utils.image_preporcess(image,
[self.input_size, self.input_size])
# 给图片加一个维度，加在最前面
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = self.sess.run(
[self.pred_sbbox, self.pred_mbbox, self.pred_lbbox],
feed_dict={
self.input_data: image_data,
self.trainable: False
})
# concatenate() 多个数组的拼接
# reshape(data, newshape)
# newshape=(-1, nums) 表示生成一个 (rows, cols)的形状，cols = nums, rows的值根据cols的值计算
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_mbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_lbbox, (-1, 5 + self.num_classes))
],
axis=0)
# 筛选出有效，在范围内的框
bboxes = utils.postprocess_boxes(pred_bbox, (org_h, org_w),
self.input_size, self.score_threshold)
# 筛选出一个最合适的框
bboxes = utils.nms(bboxes, self.iou_threshold)
return bboxes
def find_bbox(path, confidence_score=0.7, iou=0.5):
bbox_list = []
images = []
model = tf.keras.models.load_model('SavedModel/YOLOv3_model',
compile=False)
INPUT_SIZE = cfg.TEST.INPUT_SIZE
SCORE_THRESHOLD = confidence_score
IOU_THRESHOLD = iou
for filename in os.listdir(path):
full_path_image = path + '/' + filename
images.append(full_path_image)
for path_to_image in images:
image = cv2.imread(path_to_image)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_size = image.shape[:2]
image_data = utils.image_preporcess(np.copy(image),
[INPUT_SIZE, INPUT_SIZE])
image_data = image_data[np.newaxis, ...].astype(
np.float32) # (1, width, height, 3)
pred_bbox = model.predict(image_data)
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, image_size, INPUT_SIZE,
SCORE_THRESHOLD)
bboxes = utils.nms(bboxes, IOU_THRESHOLD, method='nms')
bboxes.append(path_to_image)
bbox_list.append(bboxes)
return bbox_list
def predict(self, image):
org_image = np.copy(image)
org_h, org_w, _ = org_image.shape
image_data = utils.image_preporcess(image,
[self.input_size, self.input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = self.sess.run(
[self.pred_sbbox, self.pred_mbbox, self.pred_lbbox],
feed_dict={
self.input_data: image_data,
self.trainable: False
})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_mbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_lbbox, (-1, 5 + self.num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, (org_h, org_w),
self.input_size, self.score_threshold)
bboxes = utils.nms(bboxes, self.iou_threshold)
return bboxes
def test_image(image_path, model_path):
input_size = 416
original_image = cv2.imread(image_path)
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image), [input_size, input_size])
image_data = image_data[np.newaxis, ...].astype(np.float32)
model = yolov3.build_for_test()
# 加载tf model:model.load_weights(model_path);加载darknet model: utils.load_weights(model, model_path)
utils.load_weights(model, model_path)
model.summary()
start_time = time.time()
pred_bbox = model.predict(image_data)
print('pred_bbox>>>>>>>>>>>>>>>>>', pred_bbox)
end_time = time.time()
print("time: %.2f ms" %(1000*(end_time-start_time)))
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
# 将416×416下的bbox坐标转换为原图上的坐标并删除部分无效box
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size, input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
# 构建原图和bbox画出坐标框
image = utils.draw_bbox(original_image, bboxes)
image = Image.fromarray(image)
image.show()
def predict(self, image):
org_image = np.copy(image)
org_h, org_w, _ = org_image.shape
image_data = utils.image_preporcess(image,
[self.input_size, self.input_size])
image_data = image_data[np.newaxis, ...]
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
pred_sbbox, pred_mbbox, pred_lbbox, conv_sbbox, conv_mbbox, conv_lbbox = self.sess.run(
[
self.pred_sbbox, self.pred_mbbox, self.pred_lbbox,
self.conv_sbbox, self.conv_mbbox, self.conv_lbbox
],
feed_dict={
self.input_data: image_data,
self.trainable: False
},
options=options,
run_metadata=run_metadata)
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 9 + self.num_classes)),
np.reshape(pred_mbbox, (-1, 9 + self.num_classes)),
np.reshape(pred_lbbox, (-1, 9 + self.num_classes))
],
axis=0)
bboxes1 = utils.postprocess_boxes(pred_bbox, (org_h, org_w),
self.input_size,
self.score_threshold)
bboxes = utils.nms(bboxes1, self.iou_threshold)
return bboxes
def predict(self, image_path):
original_image = cv2.imread(image_path)
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[self.input_size, self.input_size])
image_data = image_data[np.newaxis, ...]
with tf.Session(graph=self.graph) as sess:
pred_sbbox, pred_mbbox, pred_lbbox = self.sess.run(
[
self.return_tensors[1], self.return_tensors[2],
self.return_tensors[3]
],
feed_dict={self.return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_mbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_lbbox, (-1, 5 + self.num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
self.input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image, label = utils.draw_bbox(original_image, bboxes)
# image = Image.fromarray(image)
# print(image)
cv2.imwrite('static/images/test.jpg', image)
# image.show()
return label
def batch_bboxes(model, frames):
all_image_data = [None] * len(frames)
sizes = [None] * len(frames)
bbbb = [[], []]
for i, frame in enumerate(frames):
#move frame to GPU
all_image_data[i] = frame_to_gpu(frame)
sizes[i] = frame.shape[:2]
stacked = tf.stack(all_image_data, axis=1)
trimmed = tf.squeeze(stacked)
# print(trimmed.shape)
# dataset = tf.data.Dataset.from_tensor_slices(stacked)
# dataset = dataset.batch(2)
# for im_data in dataset.as_numpy_iterator():
pred_bbox = model.predict(trimmed)
print('aww yeah')
pred_bbox = utils.postprocess_bbbox(pred_bbox, ANCHORS, STRIDES, XYSCALE)
all_bboxes, probs, classes, image_nums = utils.postprocess_boxes(
pred_bbox, sizes[0], INPUT_SIZE, 0.25) #.25
bboxes = utils.filter_people(all_bboxes, probs, classes)
bboxes = utils.nms(bboxes, 0.213, method='nms')
bboxes = np.array(bboxes)
# bboxes2 = utils.nms(bboxes2, 0.213, method='nms')
# bboxes2 = np.array(bboxes2)
print('frame1')
def predicate(original_image, pb_file):
if pb_file == None:
pb_file = DEFAULT_PB
return_elements = ["input/input_data:0", "pred_sbbox/concat_2:0", "pred_mbbox/concat_2:0", "pred_lbbox/concat_2:0"]
num_classes = 66
input_size = 608
graph = tf.Graph()
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image), [input_size, input_size])
image_data = image_data[np.newaxis, ...]
return_tensors = utils.read_pb_return_tensors(graph, pb_file, return_elements)
with tf.compat.v1.Session(graph=graph) as sess:
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={ return_tensors[0]: image_data})
pred_bbox = np.concatenate([np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))], axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size, input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
return bboxes, original_image
def predict(self, input_rgb,input_lwir):
org_image_rgb = np.copy(input_rgb)
org_image_lwir = np.copy(input_lwir)
org_h, org_w, _ = org_image_rgb.shape
img_rgb,img_lwir = utils.image_preporcess(org_image_rgb,org_image_lwir, [self.input_size, self.input_size])
img_rgb = img_rgb[np.newaxis, ...]
img_lwir = img_lwir[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = self.sess.run(
[self.pred_sbbox, self.pred_mbbox, self.pred_lbbox],
feed_dict={
self.input_rgb: img_rgb,
self.input_lwir: img_lwir,
self.trainable: False
}
)
pred_bbox = np.concatenate([np.reshape(pred_sbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_mbbox, (-1, 5 + self.num_classes)),
np.reshape(pred_lbbox, (-1, 5 + self.num_classes))], axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, (org_h, org_w), self.input_size, self.score_threshold)
bboxes = utils.nms(bboxes, self.iou_threshold)
return bboxes
def calculate(self, frame):
start = int(round(time.time() * 1000))
# Image to be processed
original_image = frame
# Read class names
class_names = {}
with open(cfg.YOLO.CLASSES, 'r') as data:
for ID, name in enumerate(data):
class_names[ID] = name.strip('\n')
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[self.input_size, self.input_size])
image_data = image_data[np.newaxis, ...].astype(np.float32)
pred_bbox = self.model.predict(image_data)
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
self.input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
# We have our objects detected and boxed, lets move the class name into a list
objects_detected = []
for x0, y0, x1, y1, prob, class_id in bboxes:
objects_detected.append(class_names[class_id])
# Lets show the user a nice picture - should be erased in production
#image = utils.draw_bbox(original_image, bboxes)
#image = Image.fromarray(image)
#image.show()
exec_time = int(round(time.time() * 1000)) - start
print(f"Objects Detected: {objects_detected}")
return objects_detected, exec_time
def detect_img(fim, start_h, start_w, fpredict_bboxes):
img_size = fim.shape[:2]
image_data = utils.image_preporcess(np.copy(fim),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, img_size,
input_size, score_thresh)
bboxes = utils.nms(bboxes, iou_type, iou_thresh, method='nms')
for i, bbox in enumerate(bboxes):
# bboxes: [x_min, y_min, x_max, y_max, probability, cls_id] format coordinates
coor = np.array(bbox[:4], dtype=np.int32)
bbox[0] = int(start_w + coor[0])
bbox[1] = int(start_h + coor[1])
bbox[2] = int(start_w + coor[2])
bbox[3] = int(start_h + coor[3])
fpredict_bboxes.append(bbox)
def alone_bbox(image_path, save_path):
class_list = ["bench", "roadblock", "babycar", "wheelchair"]
return_elements = [
"input/input_data:0", "pred_sbbox/concat_2:0", "pred_mbbox/concat_2:0",
"pred_lbbox/concat_2:0"
]
input_size = cfg.TEST.INPUT_SIZE
graph = tf.Graph()
save_file_name = osp.basename(image_path).split('.')[0] + '.txt'
original_image = cv2.imread(image_path)
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
with open(osp.join(save_path, save_file_name), 'a') as fp:
name_dict_list = []
bboxes_list = []
for newlabel in class_list:
pb_file = "./yolov3_{}.pb".format(newlabel)
name_dict = utils.read_class_names(
"./data/classes/{}.names".format(newlabel))
name_dict_list.append(name_dict)
num_classes = len(name_dict)
return_tensors = utils.read_pb_return_tensors(
graph, pb_file, return_elements)
with tf.Session(graph=graph) as sess:
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, 0.9)
bboxes = utils.nms(bboxes, 0.90, method='nms')
bboxes_list.append(bboxes)
o = str(' ' + str(0))
for bbox in bboxes:
min_x, min_y, max_x, max_y = [
str(int(bbox[i])) for i in range(4)
]
label_name = name_dict[int(bbox[5])]
fp.writelines(label_name + o + o + o + ' ' + min_x + ' ' +
min_y + ' ' + max_x + ' ' + max_y + o + o + o +
o + o + o + o)
fp.writelines('\n')
# for i,bboxes_ in enumerate(bboxes_list):
# image = utils.draw_bbox(original_image, bboxes_, name_dict_list[i])
# image = Image.fromarray(image)
# image.show()
return bboxes_list, name_dict_list
def predict(image_path):
original_image = cv2.imread(image_path) # 读取图片
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
return_tensors = utils.read_pb_return_tensors(graph, pb_file,
return_elements)
with tf.Session(graph=graph) as sess:
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, 0.35)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(original_image, bboxes)
image = Image.fromarray(image)
image.show()
image.save(output_path)
def coco_bbox(image_path, save_path, return_tensorsm, graph):
t1 = time.time()
# return_elements = ["input/input_data:0", "pred_sbbox/concat_2:0", "pred_mbbox/concat_2:0", "pred_lbbox/concat_2:0"]
input_size = cfg.TEST.INPUT_SIZE
# graph = tf.Graph()
save_file_name = osp.basename(image_path).split('.jp')[0] + '.txt'
original_image = cv2.imread(image_path)
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
# pb_file = cfg.TEST.PB_FILE
name_dict = utils.read_class_names(cfg.TEST.CLASSES)
num_classes = len(name_dict)
choose_classes = cfg.TEST.CHOOSE_CLASSES
with open(osp.join(save_path, save_file_name), 'w') as fp:
# return_tensors = utils.read_pb_return_tensors(graph, pb_file, return_elements)
with tf.compat.v1.Session(graph=graph) as sess:
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
# t2 = time.time()
# print("time2: ", t2 - t1)
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, 0.45)
bboxes = utils.nms(bboxes, 0.35, method='nms')
o = str(' ' + str(0))
# with open(osp.join(save_path, save_file_name), 'w') as fp:
for bbox in bboxes:
min_x, min_y, max_x, max_y = [str(int(bbox[i])) for i in range(4)]
score = str(round(bbox[4], 2))
label_name = name_dict[int(bbox[5])]
if label_name == 'person':
label_name_choose = 'Pedestrian'
elif label_name == 'bicycle' or label_name == 'motorbike':
label_name_choose = 'Cyclist'
elif label_name == 'car' or label_name == 'bus' or label_name == 'truck':
label_name_choose = 'Car'
else:
label_name_choose = 'DontCare'
if label_name in choose_classes:
fp.writelines(label_name + o + o + o + ' ' + min_x + ' ' +
min_y + ' ' + max_x + ' ' + max_y + o + o + o +
o + o + o + o + " " + score)
fp.writelines('\n')
# image = utils.draw_bbox(original_image, bboxes,name_dict)
# image = Image.fromarray(image)
# image.show()
return original_image, bboxes, name_dict
def predict_logo_boxes(image_path, input_size, num_classes, original_image, return_tensors, sess):
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image), [input_size, input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_yolo_bbox = np.concatenate([np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))], axis=0)
post_yolo_boxes = utils.postprocess_boxes(pred_yolo_bbox, original_image_size, input_size, 0.1)
post_yolo_boxes = utils.nms(post_yolo_boxes, 0.2, method='nms')
post_yolo_boxes = ipm.logo_filter(post_yolo_boxes)
ocr_boxes = ipm.text_detection(image_path)
# cost_match_matrix refers to the distance pairs between the yolo boxes and ocr boxes
cost_match_matrix = ipm.cost_calculating(post_yolo_boxes, ocr_boxes)
row_ind, col_ind = linear_sum_assignment(cost_match_matrix)
ocr_boxes = [ocr_boxes[t] for t in col_ind]
yolo_boxes = [post_yolo_boxes[t] for t in row_ind]
distance = cost_match_matrix[row_ind, col_ind]
mask = [d < 100 for d in distance]
yolo_boxes = np.array(yolo_boxes)
ocr_boxes = np.array(ocr_boxes)
yolo_boxes = yolo_boxes[mask]
yolo_boxes = yolo_boxes.tolist()
ocr_boxes = ocr_boxes[mask]
ocr_boxes = ocr_boxes.tolist()
return ocr_boxes, post_yolo_boxes
def on_run(image):
image_data = utils.image_preporcess(np.copy(image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...].astype(np.float32)
pred_bbox = model.predict(image_data)
# sys.stdout.write(f"[yolo_detect] pred_bbox {pred_bbox}")
# sys.stdout.write(f"[yolo_detect] pred_bbox[0].shape {pred_bbox[0].shape}")
# sys.stdout.write(f"[yolo_detect] anchors {anchors}")
# sys.stdout.flush()
if model_name == "yolov3":
pred_bbox = utils.postprocess_bbbox(pred_bbox, anchors, STRIDES)
else:
pred_bbox = utils.postprocess_bbbox(pred_bbox, anchors, STRIDES,
XYSCALE)
bboxes = utils.postprocess_boxes(pred_bbox, image.shape[:-1], input_size,
conf_threshold)
bboxes = utils.nms(bboxes, iou_threshold, method='nms')
# bboxes[[xmin, ymin, xmax, ymax, score, class]]
# sys.stdout.write(f"[yolov4 detect] bboxes {bboxes}")
# sys.stdout.flush()
return {'bboxes': np.array(bboxes)}
def process(image_path, targetFolder, sess, return_tensors):
head, tail = os.path.split(image_path)
localFileName = tail
original_image = cv2.imread(image_path)
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, score_threshold)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(original_image, bboxes)
image = Image.fromarray(image)
exportName = "out_" + localFileName
filepath = targetFolder + "/" + exportName
image.save(filepath)
def predict(self, images):
org_h = [0 for i in range(self.batch_size)]
org_w = [0 for i in range(self.batch_size)]
for i in range(self.batch_size):
org_h[i], org_w[i], _ = images[i].shape
image_data = utils.images_preporcess(
images, [self.input_size, self.input_size])
start = time.time()
pred_sbbox, pred_mbbox, pred_lbbox = self.sess.run(
[self.pred_sbbox, self.pred_mbbox, self.pred_lbbox],
feed_dict={
self.input_data: image_data,
})
end = time.time()
print("inference time exclude postprocess is: ", (end - start) * 1000)
batch_bboxes = []
for idx in range(self.batch_size):
pred_bbox = np.concatenate([
np.reshape(pred_sbbox[idx], (-1, 5 + self.num_classes)),
np.reshape(pred_mbbox[idx], (-1, 5 + self.num_classes)),
np.reshape(pred_lbbox[idx], (-1, 5 + self.num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox,
(org_h[idx], org_w[idx]),
self.input_size,
self.score_threshold)
batch_bboxes.append(utils.nms(bboxes, self.iou_threshold))
end_ = time.time()
print("inference time include postprocess is: ", (end_ - start) * 1000)
return batch_bboxes, (end - start)
def detect_images(model,
image_path,
box=None,
output_path="",
id=0,
write_file=True,
show=False):
"""Object classification of the given image.
Run the yolo model on the given image. With post process including nms.
Save the output image to file or show the image if specified.
Args:
model: The yolo model to be used.
image_path: path to the image.
box: bounding box coordinates. Should be a list like: [x1, y1, x2, y2].
output_path: path to write the output image.
id: index of bounding box for a given frame.
show: whether to show the image for display.
"""
original_image = cv2.imread(image_path)
if box:
original_image = original_image[box[1]:box[3], box[0]:box[2]]
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data, old_image_size, new_image_size = utils.image_preprocess(
np.copy(original_image))
image_data = image_data[np.newaxis, ...].astype(np.float32)
# pred_bbox = model.predict(image_data)
pred_bbox = model.predict_on_batch(image_data)
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, old_image_size, new_image_size,
0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(original_image, bboxes)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if id:
i = output_path.rfind('.')
output_path = output_path[:i] + '_' + str(id) + output_path[i:]
if output_path != '' and write_file:
i = output_path.rfind('/')
output_directory = output_path[:i]
if not os.path.exists(output_directory):
os.makedirs(output_directory)
cv2.imwrite(output_path, image)
if show:
# Show the image
cv2.imshow("predicted image", image)
# Load and hold the image
cv2.waitKey(0)
# To close the window after the required kill value was provided
cv2.destroyAllWindows()
def mul_image(watch_dir="./docs/images", output_path='./output'):
imageDir = os.path.abspath(watch_dir)
imageList = glob.glob(os.path.join(imageDir, '*.jpg'))
# print(imageList)
graph = tf.Graph()
pb_file = "./yolov3_coco_v3.pb"
return_elements = [
"input/input_data:0", "pred_sbbox/concat_2:0", "pred_mbbox/concat_2:0",
"pred_lbbox/concat_2:0"
]
return_tensors = utils.read_pb_return_tensors(graph, pb_file,
return_elements)
with tf.Session(graph=graph) as sess:
for item in imageList:
image_path = item
# print('item',item)
end = "/"
name = item[item.rfind(end):]
# print(name)
num_classes = 80
input_size = 608
out = output_path + name
original_image = cv2.imread(image_path)
# original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, 0.45)
# print('bboxes:',bboxes)
# bboxes: [[301.13088989 118.44700623 346.95623779 172.39486694 0.97461057 0]...]
bboxes = utils.nms(bboxes, 0.45, method='nms')
# print('bboxes:',bboxes)
# bboxes: [array([105.31238556, 54.51167679, 282.53552246, 147.27146912, 0.99279714, 0. ])]
image = utils.draw_bbox(original_image, bboxes)
cv2.imwrite(out, image)
def vehicle_detection(image):
protocol = ProtocolType.from_str('http')
input_name = 'input/input_data'
sboxes = 'pred_sbbox/concat_2'
mboxes = 'pred_mbbox/concat_2'
lboxes = 'pred_lbbox/concat_2'
model_name = 'vehicle-detector'
ctx = InferContext('localhost:8000', protocol, model_name, -1, False)
ori_size = image.shape[:2]
image_preprocess = utils.image_preporcess(image, [416, 416])
image_data = []
image_data.append(image_preprocess)
result = []
image_idx = 0
request_ids = []
batch_size = 4
last_request = False
input_batch = []
while not last_request:
input_batch = []
for idx in range(batch_size):
input_batch.append(image_data[image_idx])
image_idx = (image_idx + 1) % len(image_data)
if image_idx == 0:
last_request = True
batch_size = len(input_batch)
break
request_ids.append(
ctx.async_run({input_name: input_batch}, {
sboxes: (InferContext.ResultFormat.RAW),
mboxes: (InferContext.ResultFormat.RAW),
lboxes: (InferContext.ResultFormat.RAW)
}, batch_size))
raw_results = []
num_classes = 80
# # For async, retrieve results according to the send order
for request_id in request_ids:
raw_results.append(ctx.get_async_run_results(request_id, True))
for ix in range(len(raw_results)):
pred_sbbox, pred_mbbox, pred_lbbox = raw_results[ix][
'pred_sbbox/concat_2'], raw_results[ix][
'pred_mbbox/concat_2'], raw_results[ix]['pred_lbbox/concat_2']
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, ori_size, 416, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
return bboxes
def video_without_saving():
classes = utils.read_class_names(cfg.YOLO.CLASSES)
num_classes = len(classes)
return_elements = [
"input/input_data:0", "pred_sbbox/concat_2:0", "pred_mbbox/concat_2:0",
"pred_lbbox/concat_2:0"
]
pb_file = "./yolov3_coco.pb"
video_path = "docs/images/racoon.mp4"
video_path = 0
input_size = 416
graph = tf.Graph()
return_tensors = utils.read_pb_return_tensors(graph, pb_file,
return_elements)
with tf.Session(graph=graph) as sess:
vid = cv2.VideoCapture(video_path)
while True:
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
image = Image.fromarray(frame)
else:
raise ValueError("No image!")
frame_size = frame.shape[:2]
image_data = utils.image_preporcess(np.copy(frame),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
prev_time = time.time()
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, frame_size, input_size,
0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(frame, bboxes)
curr_time = time.time()
exec_time = curr_time - prev_time
result = np.asarray(image)
info = "time: %.2f ms" % (1000 * exec_time)
cv2.namedWindow("result", cv2.WINDOW_AUTOSIZE)
result = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
cv2.imshow("result", result)
if cv2.waitKey(1) & 0xFF == ord('q'): break
def gen():
for i in range(5):
count = 0
cap = cv2.VideoCapture("full_video.mp4")
start_frame_number = 1000
cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame_number)
while True:
cap.set(cv2.CAP_PROP_POS_MSEC, (count * 125))
ret, frame = cap.read()
count += 1
if not ret:
break
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
image = Image.fromarray(frame)
frame_size = frame.shape[:2]
image_data = utils.image_preporcess(np.copy(frame),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
image_data = image_data.astype(np.float32)
prev_time = time.time()
request.inputs['input'].CopyFrom(
tf.contrib.util.make_tensor_proto(
image_data, shape=[1, input_size, input_size, 3]))
result_future = stub.Predict.future(request, 10.25)
pred_sbbox = np.asarray(
result_future.result().outputs['pred_sbbox'].float_val)
pred_mbbox = np.asarray(
result_future.result().outputs['pred_mbbox'].float_val)
pred_lbbox = np.asarray(
result_future.result().outputs['pred_lbbox'].float_val)
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, frame_size, input_size,
0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(frame, bboxes)
curr_time = time.time()
exec_time = curr_time - prev_time
print(exec_time)
result = np.asarray(image)
frame = cv2.cvtColor(result, cv2.COLOR_RGB2BGR)
#result = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
#format should be bgr
ret, jpeg = cv2.imencode('.jpg', frame)
frame = jpeg.tobytes()
yield (b'--frame\r\n'
b'Content-Type: image/jpeg\r\n\r\n' + frame + b'\r\n\r\n')
def show(pred_bbox):
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(original_image, bboxes, show_label=True)
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
cv2.imshow('sample', image)
cv2.waitKey(0)
def detection(vid):
with tf.Session(graph=graph) as sess:
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
image = Image.fromarray(frame)
else:
raise ValueError("No image!")
frame_size = frame.shape[:2]
image_data = utils.image_preporcess(np.copy(frame),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
prev_time = time.time()
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, frame_size, input_size,
0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image, detected = utils.draw_bbox(frame, bboxes)
detected = np.asarray(detected)
# print("------- frame i ---------")
class_count = []
for i in range(len(obj_classes)): # 80
obj_count = 0
for j in range(len(detected)):
if int(detected[j][5]) == i: obj_count += 1
class_count = np.append(class_count, obj_count)
curr_time = time.time()
exec_time = curr_time - prev_time
result = np.asarray(image)
info = "time: %.2f ms" % (1000 * exec_time)
# cv2.namedWindow("result", cv2.WINDOW_AUTOSIZE)
result = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
return result, class_count
def get_object(video_url, threshold=0.45):
input_layer = tf.keras.layers.Input([input_size, input_size, 3])
feature_maps = YOLOv3(input_layer)
bbox_tensors = []
for i, fm in enumerate(feature_maps):
bbox_tensor = decode(fm, i)
bbox_tensors.append(bbox_tensor)
model = tf.keras.Model(input_layer, bbox_tensors)
utils.load_weights(model, "yolov3_union_10000.weights")
model.summary()
vid = cv2.VideoCapture(video_url)
while True:
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
else:
return "No image"
frame_size = frame.shape[:2]
image_data = utils.image_preporcess(np.copy(frame),
[input_size, input_size])
image_data = image_data[np.newaxis, ...].astype(np.float32)
prev_time = time.time()
pred_bbox = model.predict_on_batch(image_data)
curr_time = time.time()
exec_time = curr_time - prev_time
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, frame_size, input_size,
0.3)
bboxes = utils.nms(bboxes, threshold, method='nms')
image = utils.draw_bbox(frame, bboxes)
result = np.asarray(image)
info = "time: %.2f ms" % (1000 * exec_time)
cv2.putText(result,
text=info,
org=(50, 70),
fontFace=cv2.FONT_HERSHEY_SIMPLEX,
fontScale=1,
color=(255, 0, 0),
thickness=2)
result = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
retval, buffer = cv2.imencode(".jpeg", image)
yield ((b'--frame\r\n'
b'Content-Type: image/jpeg\r\n\r\n' + buffer.tobytes() +
b'\r\n'))
def export_yolo_video(video_path, output_path):
vid = cv2.VideoCapture(video_path)
out = cv2.VideoWriter(
str(Path(output_path,
Path(video_path).stem + '_output.mp4')),
cv2.VideoWriter_fourcc(*'mp4v'),
float(vid.get(cv2.CAP_PROP_FPS)),
(round(vid.get(cv2.CAP_PROP_FRAME_WIDTH)),
round(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))),
)
outputs = []
print(Path(output_path, Path(video_path).stem + '_output.mp4'))
print(Path(output_path, Path(video_path).stem + '_output.npz'))
max_frames = vid.get(cv2.CAP_PROP_FRAME_COUNT)
with tf.Session(graph=graph) as sess:
for _ in trange(int(max_frames)):
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
else:
raise ValueError("No image!")
frame_size = frame.shape[:2]
image_data = utils.image_preporcess(np.copy(frame),
[input_size, input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, frame_size, input_size,
0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
outputs.append(bboxes)
image = utils.draw_bbox(frame, bboxes)
result = np.asarray(image)
result = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
out.write(result)
vid.release()
out.release()
np.savez(str(Path(output_path,
Path(video_path).stem + '_output.npz')), outputs)
def do_video(self):
vid = cv2.VideoCapture(self.video_path)
while True:
# frame 是 RGB 颜色空间
return_value, frame = vid.read()
if return_value:
# utils.image_preporcess 这个方法里面有 cv2.COLOR_BGR2RGB 方法
# 如果自己写的模型，可以调一下，也许不需要这里
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
pass
else:
raise ValueError("No image!")
pass
frame_size = frame.shape[:2]
# 之前训练的时候，转了一次颜色空间
image_data = utils.image_preporcess(
np.copy(frame), [self.input_size, self.input_size])
image_data = image_data[np.newaxis, ...]
pred_start_time = datetime.now()
pred_sbbox, pred_mbbox, pred_lbbox = self.sess.run(
[
self.return_tensors[1], self.return_tensors[2],
self.return_tensors[3]
],
feed_dict={self.return_tensors[0]: image_data})
pred_bbox = np.concatenate([
np.reshape(pred_sbbox, (-1, 5 + self.class_name_len)),
np.reshape(pred_mbbox, (-1, 5 + self.class_name_len)),
np.reshape(pred_lbbox, (-1, 5 + self.class_name_len))
],
axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, frame_size,
self.input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
image = utils.draw_bbox(frame, bboxes)
pred_end_time = datetime.now()
print("一帧耗时: {}".format(pred_end_time - pred_start_time))
cv2.namedWindow("result", cv2.WINDOW_AUTOSIZE)
result = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
cv2.imshow("result", result)
# 退出按键
if cv2.waitKey(1) & 0xFF == ord('q'):
break
pass
pass
def detect(image_path):
#original_image=Image.open(BytesIO(original_image)).convert("RGBA")
original_image = cv2.imread(
image_path
) #you can and should replace this line to receive the image directly (not from a file)
#original_image = base64.b64decode(dec_image)
# Read class names
class_names = {}
with open(cfg.YOLO.CLASSES, 'r') as data:
for ID, name in enumerate(data):
class_names[ID] = name.strip('\n')
# Setup tensorflow, keras and YOLOv3
input_size = 416
input_layer = tf.keras.layers.Input([input_size, input_size, 3])
feature_maps = YOLOv3(input_layer)
original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image),
[input_size, input_size])
image_data = image_data[np.newaxis, ...].astype(np.float32)
bbox_tensors = []
for i, fm in enumerate(feature_maps):
bbox_tensor = decode(fm, i)
bbox_tensors.append(bbox_tensor)
model = tf.keras.Model(input_layer, bbox_tensors)
load_weights(model, "./yolov3.weights")
pred_bbox = model.predict(image_data)
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size,
input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
# We have our objects detected and boxed, lets move the class name into a list
objects_detected = []
for x0, y0, x1, y1, prob, class_id in bboxes:
objects_detected.append(class_names[class_id])
# Lets show the user a nice picture - should be erased in production
#image = utils.draw_bbox(original_image, bboxes)
#image = Image.fromarray(image)
#image.show()
return objects_detected
