def test_warp_nonmatching_resolutions():
previous, current = (np.random.random((128, 128, 1)) for _ in range(2))
flow = numerical_optical_flow(previous, current)
features = np.random.random((32, 32, 16))
warped = warp(features, flow)
scaled_before_warp = cv2.resize(warp(cv2.resize(features, (128, 128)), flow), (32, 32))
relative_error = np.abs((warped - scaled_before_warp) / warped)
assert np.mean(relative_error) < .08
def test_optical_flow_reuse_destination():
resolution = (128, 128)
previous, current = (np.random.randint(256, size=(*resolution, 1), dtype=np.uint8) for _ in range(2))
flow = np.empty((*resolution, 2), dtype=np.float32)
numerical_optical_flow(previous, current, destination=flow)
assert (flow == numerical_optical_flow(previous, current)).all()
warped = np.empty(resolution, dtype=np.float32)
warp(previous, flow, destination=warped)
assert (warped == warp(previous, flow)).all()
def apply_optical_flow(capture):
def get_frame():
valid, frame = capture.read()
return frame if valid else None
frame = get_frame()
if frame is None:
return
previous = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
flow = np.empty((*frame.shape[:2], 2), dtype=np.float32)
frame = get_frame()
while frame is not None:
current = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
numerical_optical_flow(previous, current, destination=flow)
warp(previous, flow, destination=previous)
cv2.imshow('Warped', previous)
cv2.waitKey(1)
frame = get_frame()
def test_warp():
previous, current = (np.random.random((128, 128, 1)) for _ in range(2))
flow = numerical_optical_flow(previous, current)
assert flow.shape == (128, 128, 2)
assert warp(previous, flow).shape == (128, 128)
def test_optical_flow_normalized():
# Should return the same results for normalized and [0, 255] images
previous, current = (np.random.randint(256, size=(128, 128, 1), dtype=np.uint8) for _ in range(2))
flow = numerical_optical_flow(previous, current)
assert (flow == numerical_optical_flow(previous / 255, current / 255)).all()
assert np.allclose(warp(previous, flow) / 255, warp(previous / 255, flow))
def test_warp_multiple_filters():
previous, current = (np.random.random((128, 128, 1)) for _ in range(2))
flow = numerical_optical_flow(previous, current)
assert warp(np.random.random((32, 32, 16)), flow).shape == (32, 32, 16)
def warp_features(l_input_tm1, l_input, features_tm1):
# TODO Replace with a optical flow network and optimize globally with the rest of the model
# TODO Reuse destination for performance
flow = numerical_optical_flow(l_input_tm1, l_input)
return warp(features_tm1, flow)