"""
Implementation of global motion estimators.
"""
import numpy as np
import os
import scipy.ndimage
import scipy.spatial
from ..utils import *
def _hausdorff_distance(E_1, E_2):
# binary structure
diamond = np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])
# extract only 1-pixel border line of objects
E_1_per = E_1^scipy.ndimage.binary_erosion(E_1, structure=diamond)
E_2_per = E_2^scipy.ndimage.binary_erosion(E_2, structure=diamond)
A = scipy.ndimage.distance_transform_edt(~E_2_per)[E_1_per].max()
B = scipy.ndimage.distance_transform_edt(~E_1_per)[E_2_per].max()
return np.max((A, B))
[docs]
def globalEdgeMotion(frame1, frame2, r=6, method='hamming'):
"""Global motion estimation using edge features
Given two frames, find a robust global translation vector
found using edge information. Frames must be luminance, RGB, or edge masks.
Parameters
----------
frame1 : ndarray
first input frame. Edge mask if dtype=bool, else luminance or rgb image. shape (1, M, N, C), (1, M, N), (M, N, C) or (M, N)
frame2 : ndarray
second input frame. Edge mask if dtype=bool, else luminance or rgb image. shape (1, M, N, C), (1, M, N), (M, N, C) or (M, N)
r : int
Search radius for measuring correspondences.
method : string
"hamming" --> use Hamming distance when measuring edge correspondence distances. The distance used in the census transform. [#f1]_
"hausdorff" --> use Hausdorff distance when measuring edge correspondence distances. [#f2]_
Returns
----------
globalMotionVector : list of int, length 2
The motion to minimize edge distances by moving frame2 with respect
to frame1. Returns ``[0, 0]`` when either frame contains no edges
(no edge correspondence to minimize), so a blank/edgeless frame
yields "no detected motion" rather than a spurious or crashing
result.
References
----------
.. [#f1] Ramin Zabih and John Woodfill. Non-parametric local transforms for computing visual correspondence. Computer Vision-ECCV, 151-158, 1994.
.. [#f2] Kevin Mai, Ramin Zabih, and Justin Miller. Feature-based algorithms for detecting and classifying scene breaks. Cornell University, 1995.
"""
frame1 = vshape(frame1)
frame2 = vshape(frame2)
if frame1.shape != frame2.shape:
raise ValueError("frame1 and frame2 must have the same shape; got %s vs %s" % (frame1.shape, frame2.shape))
T, M, N, C = frame1.shape
if C == 3:
frame1 = rgb2gray(frame1)
frame2 = rgb2gray(frame2)
elif C != 1:
raise ValueError("called with frames having %d channels. Please supply only the luminance channel or RGB images." % (C,))
# if type bool, then these are edge maps. No need to convert them, but
# still squeeze to 2D so the roll/morphology ops below see (M, N) like
# the canny() path produces (vshape made them 4D).
if frame1.dtype != bool:
E_1 = canny(frame1.squeeze())
else:
E_1 = frame1.squeeze()
if frame2.dtype != bool:
E_2 = canny(frame2.squeeze())
else:
E_2 = frame2.squeeze()
# If either frame has no edges, there is no edge correspondence to
# minimize: every displacement ties (hamming would argmin to the first,
# i.e. (-r, -r)) and the hausdorff distance does an empty-array reduction
# and crashes. Report no detectable motion.
if not E_1.any() or not E_2.any():
return [0, 0]
distances = []
displacements = []
for dx in range(-r, r+1, 1):
for dy in range(-r, r+1, 1):
cimage = np.roll(E_2, dx, axis=0)
cimage = np.roll(cimage, dy, axis=1)
# smallest distance between a point of points found in cimage
if method == 'hamming':
distance = scipy.spatial.distance.hamming(np.ravel(cimage), np.ravel(E_1))
elif method == 'hausdorff':
# cimage is the shifted frame2 edge map; measure its distance
# to frame1's edges (E_1), matching the hamming branch.
distance = _hausdorff_distance(cimage, E_1)
else:
raise NotImplementedError(
"Unknown method %r; expected 'hamming' or 'hausdorff'." % (method,)
)
# compute # of bit flip distance
distances.append(distance)
displacements.append([dx, dy])
idx = np.argmin(distances)
return displacements[idx]
