def draw_result(dev_idx, det_res, captured_frames):
#for multiple detection
for res in det_res:
x1 = int(res[0])
y1 = int(res[1])
x2 = int(res[2])
y2 = int(res[3])
class_num = res[5]
score = res[4]
# print(x1,x2,class_num,score)
if (class_num==0):
captured_frames[0] = cv2.rectangle(captured_frames[0], (x1, y1), (x2, y2), (0, 0, 255), 3)
# print("score of person: ", score)
else:
captured_frames[0] = cv2.rectangle(captured_frames[0], (x1, y1), (x2, y2), (255, 0, 0), 3)
# print("score of others: ", score)
cv2.imshow('detection', captured_frames[0])
del captured_frames[0]
key = cv2.waitKey(1)
if key == ord('q'):
kdp_wrapper.kdp_exit_dme(dev_idx)
sys.exit()
return
def user_test_single_dme(dev_idx, loop):
"""Test single dme."""
# load model into Kneron device
model_path = "../test_images/dme_ssd_fd"
is_raw_ouput = True
kdp_wrapper.kdp_dme_load_ssd_model(dev_idx, model_path, is_raw_ouput)
image_source_h = 480
image_source_w = 640
image_size = image_source_w * image_source_h * 2
frames = []
app_id = 0 # if app_id is 0, output raw data for kdp_wrapper.kdp_dme_inference
# the parameters for postprocess
anchor_path = './examples/fdssd/models/anchor_face.npy'
model_input_shape = (200, 200)
score_thres = 0.5
nms_thres = 0.35
only_max = False
# Setup video capture device.
capture = kdp_wrapper.setup_capture(0, image_source_w, image_source_h)
if capture is None:
return -1
while (loop):
raw_res = kdp_wrapper.kdp_dme_inference(dev_idx, app_id, capture, image_size, frames)
det_res = postprocess_(raw_res, anchor_path, model_input_shape,
image_source_w, image_source_h, score_thres, only_max, nms_thres)
draw_result(dev_idx, det_res, frames)
loop -= 1
kdp_wrapper.kdp_exit_dme(dev_idx)
def user_test_single_dme(dev_idx, loop):
"""Test single dme."""
# load model into Kneron device
model_path = "../test_images/dme_yolo_224"
kdp_wrapper.kdp_dme_load_yolo_model(dev_idx, model_path)
image_source_h = 480
image_source_w = 640
image_size = image_source_w * image_source_h * 2
frames = []
app_id = constants.APP_TINY_YOLO3
# Setup video capture device.
capture = kdp_wrapper.setup_capture(0, image_source_w, image_source_h)
if capture is None:
return -1
# Send 1 image to the DME image buffers.
ret, ssid = kdp_wrapper.dme_fill_buffer(dev_idx, capture, image_size,
frames)
if ret:
return -1
kdp_wrapper.dme_pipeline_inference(dev_idx, app_id, loop, image_size,
capture, ssid, frames, handle_result)
kdp_wrapper.kdp_exit_dme(dev_idx)
def user_test_single_dme(dev_idx):
"""Test single dme."""
# load model into Kneron device
model_path = "../test_images/dme_mobilenet"
kdp_wrapper.kdp_dme_load_model(dev_idx, model_path)
#get test images ready
img_path = './data/images/cat.jpg'
img_path2 = './data/images/fox.jpg'
npraw_data = kdp_wrapper.kdp_inference(dev_idx, img_path)
# Do postprocessing with keras
preds = kdp_wrapper.softmax(npraw_data[0]).reshape(1, 1000)
top_indexes(preds, 3)
#print('\nPredicted:', decode_predictions(preds, top=3)[0])
npraw_data = kdp_wrapper.kdp_inference(dev_idx, img_path2)
# Do postprocessing with keras
preds = kdp_wrapper.softmax(npraw_data[0]).reshape(1, 1000)
top_indexes(preds, 3)
#print('\nPredicted:', decode_predictions(preds, top=3)[0])
kdp_wrapper.kdp_exit_dme(dev_idx)
def user_test_single_dme(dev_idx, loop):
"""Test single dme."""
# load model into Kneron device
model_path = "../test_images/dme_ssd_fd"
is_raw_ouput = False
kdp_wrapper.kdp_dme_load_ssd_model(dev_idx, model_path, False)
image_source_h = 480
image_source_w = 640
image_size = image_source_w * image_source_h * 2
frames = []
app_id = constants.APP_FD_LM
# Setup video capture device.
capture = kdp_wrapper.setup_capture(0, image_source_w, image_source_h)
if capture is None:
return -1
while (loop):
det_res = kdp_wrapper.kdp_dme_inference(dev_idx, app_id, capture, image_size, frames)
draw_result(dev_idx, det_res, frames)
loop -= 1
# print("Total class {}: total detection {}".format(det_res[0], det_res[1]))
# for i in range(det_res[1]):
# print("x1,y1,x2,y2:", det_res[4*i+2],det_res[4*i+3],det_res[4*i+4],det_res[4*i+5])
kdp_wrapper.kdp_exit_dme(dev_idx)
def handle_result(dev_idx, raw_res, captured_frames):
# the parameters for postprocess
anchor_path = './examples/yolo/models/anchors.txt'
class_path = './common/coco_name_lists'
model_input_shape = (224, 224)
score_thres = 0.2
nms_thres = 0.45
keep_aspect_ratio = True
image_source_h = 480
image_source_w = 640
det_res = yolo_postprocess_(raw_res, anchor_path, class_path,
image_source_h, image_source_w,
model_input_shape, score_thres, nms_thres,
keep_aspect_ratio)
#for multiple detection
for res in det_res:
x1 = int(res[0])
y1 = int(res[1])
x2 = int(res[2])
y2 = int(res[3])
class_num = res[5]
score = res[4]
# print(x1,x2,class_num,score)
if (class_num == 0):
captured_frames[0] = cv2.rectangle(captured_frames[0], (x1, y1),
(x2, y2), (0, 0, 255), 3)
# print("score of person: ", score)
else:
captured_frames[0] = cv2.rectangle(captured_frames[0], (x1, y1),
(x2, y2), (255, 0, 0), 3)
# print("score of others: ", score)
cv2.imshow('detection', captured_frames[0])
del captured_frames[0]
key = cv2.waitKey(1)
if key == ord('q'):
kdp_wrapper.kdp_exit_dme(dev_idx)
sys.exit()
return
def draw_result(dev_idx, det_res, captured_frames):
x1_0 = 0
y1_0 = 0
x2_0 = 0
y2_0 = 0
score_0 = 0
#for multiple faces
for res in det_res:
#print(type(res))
x1 = int(res[0])
y1 = int(res[1])
x2 = int(res[2]+res[0])
y2 = int(res[3]+res[1])
class_num = res[5]
score = res[4]
o_l = overlap(x1,y1,x2,y2,x1_0,y1_0,x2_0,y2_0)
if (o_l<0.6):
x1_0 = x1
y1_0 = y1
x2_0 = x2
y2_0 = y2
score_0 = score
if (class_num==2):
captured_frames[0] = cv2.rectangle(captured_frames[0], (x1, y1), (x2, y2), (0, 0, 255), 3)
#print("score of mask fd: ", score)
elif (class_num==1):
captured_frames[0] = cv2.rectangle(captured_frames[0], (x1, y1), (x2, y2), (255, 0, 0), 3)
#print("score of fd: ", score)
cv2.imshow('detection', captured_frames[0])
del captured_frames[0]
key = cv2.waitKey(1)
if key == ord('q'):
kdp_wrapper.kdp_exit_dme(dev_idx)
sys.exit()
return
def user_test_single_dme(dev_idx, loop):
"""Test single dme."""
# load model into Kneron device
model_path = "../test_images/dme_yolo_224"
kdp_wrapper.kdp_dme_load_yolo_model(dev_idx, model_path)
image_source_h = 480
image_source_w = 640
image_size = image_source_w * image_source_h * 2
frames = []
app_id = constants.APP_TINY_YOLO3
# the parameters for postprocess
anchor_path = './examples/yolo/models/anchors.txt'
class_path = './common/coco_name_lists'
model_input_shape = (224, 224)
score_thres = 0.2
nms_thres = 0.45
keep_aspect_ratio = True
# Setup video capture device.
capture = kdp_wrapper.setup_capture(0, image_source_w, image_source_h)
if capture is None:
return -1
while (loop):
raw_res = kdp_wrapper.kdp_dme_inference(dev_idx, app_id, capture, image_size, frames)
dets = yolo_postprocess_(raw_res, anchor_path, class_path, image_source_h, image_source_w, model_input_shape,
score_thres, nms_thres, keep_aspect_ratio)
# print("dets: ", dets)
draw_result(dev_idx, dets, frames)
loop -= 1
kdp_wrapper.kdp_exit_dme(dev_idx)