def predict_mask(img, net, FLAGS, DATA):
# open image
cvim = cv2.imread(img, cv2.IMREAD_UNCHANGED)
if cvim is None:
print("No image to open for ", img)
return
# predict mask from image
start = time.time()
mask = net.predict(cvim,
path=FLAGS.path + '/' + FLAGS.model,
verbose=FLAGS.verbose)
print("Prediction for img ", img, ". Elapsed: ", time.time() - start, "s")
# change to color
color_mask = util.prediction_to_color(mask, DATA["label_remap"],
DATA["color_map"])
# assess accuracy (if wanted)
if FLAGS.label is not None:
label = cv2.imread(FLAGS.label, 0)
if label is None:
print("No label to open")
quit()
net.individual_accuracy(mask, label)
cv2.imwrite(FLAGS.log + "/" + os.path.basename(img), color_mask)
if FLAGS.verbose:
# show me the image
# first, mix with image
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((im, transparent_mask, color_mask), axis=1)
util.im_tight_plt(all_img)
util.im_block()
return
def predict_mask(cvim, frame, net, FLAGS, DATA):
# predict mask from image
cvim = cv2.cvtColor(cvim, cv2.COLOR_RGB2BGR)
start = time.time()
mask = net.predict(cvim,
path=FLAGS.path + '/' + FLAGS.model,
verbose=FLAGS.verbose)
elapsed = time.time() - start
print("Prediction for frame ", frame, ". Elapsed: ", elapsed, "s")
# change to color
color_mask = util.prediction_to_color(mask, DATA["label_remap"],
DATA["color_map"])
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((im, transparent_mask), axis=1)
w, h, _ = all_img.shape
watermark = "Time: {:.3f}s, FPS: {:.3f}img/s.".format(elapsed, 1 / elapsed)
cv2.putText(all_img,
watermark,
org=(10, w - 10),
fontFace=cv2.FONT_HERSHEY_SIMPLEX,
fontScale=0.75,
color=(255, 255, 255),
thickness=2,
lineType=cv2.LINE_AA)
# write to disk
cv2.imwrite(FLAGS.log + "/mask_" + frame + ".jpg", color_mask)
cv2.imwrite(FLAGS.log + "/full_" + frame + ".jpg", all_img)
# show me the image
cv2.imshow("video", all_img.astype(np.uint8))
ch = cv2.waitKey(1)
return ch
def predict_mask(img, sess, input, output, FLAGS, DATA):
# open image
cvim = cv2.imread(img, cv2.IMREAD_UNCHANGED)
if cvim is None:
print("No image to open for ", img)
return
# predict mask from image
start = time.time()
mask = sess.run(output, feed_dict={input: [cvim]})
print("Prediction for img ", img, ". Elapsed: ", time.time() - start, "s")
# change to color
color_mask = util.prediction_to_color(mask[0, :, :], DATA["label_remap"],
DATA["color_map"])
cv2.imwrite(FLAGS.log + "/" + os.path.basename(img), color_mask)
if FLAGS.verbose:
# show me the image
# first, mix with image
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((im, transparent_mask, color_mask), axis=1)
util.im_tight_plt(all_img)
util.im_block()
return
def predict_mask(cvim, frame, net, FLAGS, DATA):
# predict mask from image
cvim = cv2.cvtColor(cvim, cv2.COLOR_RGB2BGR)
start = time.time()
mask = net.predict(cvim,
path=FLAGS.path + '/' + FLAGS.model,
verbose=FLAGS.verbose)
print("Prediction for frame ", frame, ". Elapsed: ",
time.time() - start, "s")
# change to color
color_mask = util.prediction_to_color(mask, DATA["label_remap"],
DATA["color_map"])
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((im, transparent_mask, color_mask), axis=1)
# write to disk
cv2.imwrite(FLAGS.log + "/mask_" + frame + ".jpg", color_mask)
cv2.imwrite(FLAGS.log + "/full_" + frame + ".jpg", all_img)
# show me the image
cv2.imshow("video", all_img.astype(np.uint8))
ch = cv2.waitKey(1)
return ch
def predict_mask(img, stream, d_input, d_output, context, FLAGS, DATA):
# open image
cvim = cv2.imread(img, cv2.IMREAD_UNCHANGED).astype(np.float32)
if cvim is None:
print("No image to open for ", img)
return
cvim = cv2.resize(cvim,
(DATA['img_prop']['width'], DATA['img_prop']['height']),
interpolation=cv2.INTER_LINEAR)
tcvim = np.transpose(cvim, axes=(2, 0, 1))
tcvim = tcvim.copy(order='C')
tcvim = (tcvim - 128.0) / 128.0
# Bindings provided as pointers to the GPU memory.
# PyCUDA lets us do this for memory allocations by
# casting those allocations to ints
bindings = [int(d_input), int(d_output)]
# allocate memory on the CPU to hold results after inference
output = np.empty((len(DATA['label_map']), DATA['img_prop']['height'],
DATA['img_prop']['width']),
dtype=np.float32,
order='C')
# predict mask from image
start = time.time()
cuda.memcpy_htod_async(d_input, tcvim, stream)
# execute model
context.enqueue(1, bindings, stream.handle, None)
# transfer predictions back
cuda.memcpy_dtoh_async(output, d_output, stream)
# syncronize threads
stream.synchronize()
print("Prediction for img ", img, ". Elapsed: ", time.time() - start, "s")
# mask from logits
mask = np.argmax(output, axis=0)
# change to color
color_mask = util.prediction_to_color(mask, DATA["label_remap"],
DATA["color_map"])
# save to log folder
cv2.imwrite(FLAGS.log + "/" + os.path.basename(img), color_mask)
if FLAGS.verbose:
# show me the image
# first, mix with image
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((im, transparent_mask, color_mask), axis=1)
util.im_tight_plt(all_img)
util.im_block()
return
def predict_mask(cvim, frame, stream, d_input, d_output, context, FLAGS, DATA):
# do all required transpositions
cvim = cv2.cvtColor(cvim, cv2.COLOR_RGB2BGR)
cvim = cv2.resize(cvim.astype(np.float32),
(DATA['img_prop']['width'], DATA['img_prop']['height']),
interpolation=cv2.INTER_LINEAR)
tcvim = np.transpose(cvim, axes=(2, 0, 1))
tcvim = tcvim.copy(order='C')
tcvim = (tcvim - 128.0) / 128.0
# Bindings provided as pointers to the GPU memory.
# PyCUDA lets us do this for memory allocations by
# casting those allocations to ints
bindings = [int(d_input), int(d_output)]
# allocate memory on the CPU to hold results after inference
output = np.empty((len(DATA['label_map']), DATA['img_prop']['height'],
DATA['img_prop']['width']),
dtype=np.float32,
order='C')
# predict mask from image
start = time.time()
cuda.memcpy_htod_async(d_input, tcvim, stream)
# execute model
context.enqueue(1, bindings, stream.handle, None)
# transfer predictions back
cuda.memcpy_dtoh_async(output, d_output, stream)
# syncronize threads
stream.synchronize()
print("Prediction for frame ", frame, ". Elapsed: ",
time.time() - start, "s")
# mask from logits
mask = np.argmax(output, axis=0)
# change to color
color_mask = util.prediction_to_color(mask, DATA["label_remap"],
DATA["color_map"])
# transparent
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((cvim, transparent_mask, color_mask), axis=1)
# write to disk
cv2.imwrite(FLAGS.log + "/mask_" + frame + ".jpg", color_mask)
cv2.imwrite(FLAGS.log + "/full_" + frame + ".jpg", all_img)
# show me the image
cv2.imshow("video", all_img.astype(np.uint8))
ch = cv2.waitKey(1)
return ch
def predict_mask(cvim, frame, sess, input, output, FLAGS, DATA):
# predict mask from image
cvim = cv2.cvtColor(cvim, cv2.COLOR_RGB2BGR)
start = time.time()
mask = sess.run(output, feed_dict={input: [cvim]})
print("Prediction for frame ", frame, ". Elapsed: ",
time.time() - start, "s")
# change to color
color_mask = util.prediction_to_color(mask[0, :, :], DATA["label_remap"],
DATA["color_map"])
im, transparent_mask = util.transparency(cvim, color_mask)
all_img = np.concatenate((im, transparent_mask, color_mask), axis=1)
# write to disk
cv2.imwrite(FLAGS.log + "/mask_" + frame + ".jpg", color_mask)
cv2.imwrite(FLAGS.log + "/full_" + frame + ".jpg", all_img)
# show me the image
cv2.imshow("video", all_img.astype(np.uint8))
ch = cv2.waitKey(1)
return ch
