def scaleLin(im, k):
'''Takes an image and a scale factor. Returns an image scaled using bilinear interpolation.
'''
out = io.constantIm(im.shape[0] * k, im.shape[1] * k, 0)
for y, x in imIter(out):
out[y, x] = interpolateLin(im, float(y)/k, float(x)/k, True)
return out
def scaleNN(im, k):
'''Takes an image and a scale factor. Returns an image scaled using nearest neighbor interpolation.
'''
out = io.constantIm(im.shape[0]*k, im.shape[1]*k, 0.0)
for y, x in imIter(out):
out[y,x]=im[clipY(im,round(y/k)), clipX(im,round(x/k))]
return out
def stitchH(im1, im2, H1):
'''Stitch im1 and im2 into a panorama. The resulting panorama should be in the coordinate system of im2, though
possibly extended to a larger image. That is, im2 should never appear distorted in the resulting panorama, only
possibly translated. Returns the stitched output (which may be larger than either input image).'''
H = np.linalg.inv(H1)
bbox1 = [[0,0], [im2.shape[0]-1, im2.shape[1]-1]]
bbox2 = computeTransformedBBox(im1.shape, H)
bbox = bboxUnion( bbox1, bbox2 )
trans = translate( bbox )
height = bbox[1][0] - bbox[0][0] + 1
width = bbox[1][1] - bbox[0][1] + 1
out = io.constantIm( height, width, 0.0)
ty = trans[0,2]
tx = trans[1,2]
out[ -ty:-ty+im2.shape[0], -tx:-tx+im2.shape[1] ] = im2
Htrans = np.dot(H, trans)
applyHomographyFast( im1, out, Htrans, True )
return out
def painterly(im,
texture,
N=10000,
size=50,
noise=01.3,
debug=False,
imname=''):
'''First paints at a coarse scale using all 1's for importance sampling, then paints again at size/4 scale using the sharpness map for importance sampling.'''
out = io.constantIm(im.shape[0], im.shape[1])
outCopy = None
if debug:
outCopy = out.copy()
# first pass
importance_first_pass = np.ones_like(im)
singleScalePaint(im, out, importance_first_pass, texture, size, N, noise)
if debug:
io.imwrite(out, str(imname + "PainterlyFirstPassOnly.png"))
# second pass
importance_second_pass = helper.sharpnessMap(im)
singleScalePaint(im, out, importance_second_pass, texture, size / 4, N,
noise)
if debug:
singleScalePaint(im, outCopy, importance_second_pass, texture,
size / 4, N, noise)
io.imwrite(outCopy, str(imname + "PainterlySecondPassOnly.png"))
return out
def warp(im, segmentsBefore, segmentsAfter, a=10, b=1, p=1):
'''Takes an image, a list of before segments, a list of after segments, and the parameters a,b,p (see Beier)
'''
height = im.shape[0]
width = im.shape[1]
out = io.constantIm(height, width, 0.0)
for y,x in imIter(out):
X = np.array([y,x], dtype=np.float64)
DSUM = np.array([0,0], dtype=np.float64)
weightsum = 0
for i in range(len(segmentsAfter)):
(u,v) = segmentsAfter[i].uv(X)
X_prime = segmentsBefore[i].uvtox(u,v)
D = np.subtract(X_prime, X)
dist = segmentsAfter[i].dist(X)
w = weight(segmentsAfter[i], X)
DSUM = np.add( DSUM, D*w )
weightsum += w
X_final = np.add( X, DSUM/float(weightsum) )
pixel = interpolateLin(im, X_final[0], X_final[1])
out[y,x] = pixel
return out
def convolve(im, kernel):
# Return an image filtered by kernel
shiftY, shiftX = (int(kernel.shape[0] / 2), int(kernel.shape[1] / 2))
im_out = io.constantIm(im.shape[0], im.shape[1], 0)
for y, x in imIter(im_out):
for yp, xp in imIter(kernel):
im_out[y, x] += getEdgePadded(im, y + yp - shiftY, x + xp - shiftX) * kernel[yp, xp]
return im_out
def scaleLin(im, k):
'''Takes an image and a scale factor. Returns an image scaled using bilinear interpolation.
'''
(height, width, depth) = np.shape(im)
img = io.constantIm(im.shape[0]*k, im.shape[1]*k, 0.0)
for y, x in imIter(img):
img[y][x] = interpolateLin(im, y/k, x/k, True)
return img
def warpBy1(im, src_segment, dest_segment):
h,w = im.shape[0:2]
out = imageIO.constantIm(h, w, [0, 0, 0])
for y, x in imIter(out):
dest_point = np.array([y,x])
y_p, x_p = transform(dest_point, dest_segment, src_segment)
out[y,x] = edgePaddingAccessor(y_p, x_p, im)
return interpolateLin(out, .5, .5, accessor=edgePaddingAccessor)
def boxBlur(im, k):
# Return a blured image filtered by box filter
shift = int(k / 2)
factor = 1.0 / (k ** 2)
im_out = io.constantIm(im.shape[0], im.shape[1], 0)
for y, x in imIter(im_out):
for yp, xp in [(a, b) for a in xrange(k) for b in xrange(k)]:
im_out[y, x] += getEdgePadded(im, y + yp - shift, x + xp - shift) * factor
return im_out
def interpolateLin(im, k_y, k_x, accessor=blackAccessor): #k_y, k_x better than y, x, which is confusing.
h, w = im.shape[0:2]
out = imageIO.constantIm(h, w, 0.0)
for y,x in imIter(im):
y_avr= (accessor(y-1,x,im) + accessor(y+1,x,im))/2
x_avr = (accessor(y,x-1,im) + accessor(y, x+1,im))/2
pixel_vals = (y_avr*k_y + x_avr*k_x)
out[y,x] = pixel_vals
return out
def painterly(im, texture, N=10000, size=50, noise=0.3):
'''First paints at a coarse scale using all 1's for importance sampling, then paints again at size/4 scale using the sharpness map for importance sampling.'''
importanceLow = np.ones_like(im)
outLow = io.constantIm( im.shape[0], im.shape[1], 0.0 )
singleScalePaint(im, outLow, importanceLow, texture, size, N, noise)
importanceHigh = helper.sharpnessMap(im)
singleScalePaint(im, outLow, importanceHigh, texture, size/4, N, noise)
return outLow
def scaleNN(im, k):
'''Takes an image and a scale factor. Returns an image scaled using nearest neighbor interpolation.
'''
height = im.shape[0]
width = im.shape[1]
out = io.constantIm(height*k, width*k, 0.0)
for y,x in imIter(out):
out[y,x] = im[y/k, x/k]
return out
def scaleNN(im, k):
'''Takes an image and a scale factor. Returns an image scaled using nearest neighbor interpolation.
'''
out = io.constantIm(im.shape[0] * k, im.shape[1] * k, 0)
for y, x in imIter(out):
origY = clipY(im, int(round(y/k)))
origX = clipX(im, int(round(x/k)))
out[y, x] = pix(im, origY, origX)
return out
def warpBy1(im, segmentBefore, segmentAfter):
'''Takes an image, one before segment, and one after segment.
Returns an image that has been warped according to the two segments.
'''
out = io.constantIm(im.shape[0], im.shape[1], 0)
for y, x in imIter(out):
u, v = segmentAfter.uv(np.array([y, x]))
Xprime = segmentBefore.uvtox(u, v)
out[y, x] = interpolateLin(im, Xprime[0], Xprime[1], True)
return out
def merge2images(im1, im2, H):
B=computeTransformedBBox(im1, linalg.inv(H))
B=bboxUnion(B, array([[0.0, 0.0], 1.0*array(im2.shape[0:2])]))
print 'bbox after union: ', B
T=translate(B[0, 0], B[0, 1])
out = imageIO.constantIm(B[1, 0]-B[0, 0], B[1, 1]-B[0, 1], 0)
applyHomography(im2, out, T)
applyHomography(im1, out, dot(H, T))
def bilaYUV(im, sigmaRange, sigmaY, sigmaUV):
# 6.865 only: filter YUV differently
imYUV = rgb2yuv(im)
bilateralY = bilateral(imYUV, sigmaRange, sigmaY)
bilateralUV = bilateral(imYUV, sigmaRange, sigmaUV)
im_out = io.constantIm(im.shape[0], im.shape[1], 0)
for y, x in imIter(im_out):
im_out[y, x] = np.array([bilateralY[y, x, 0], bilateralUV[y, x, 1], bilateralUV[y, x, 2]])
return yuv2rgb(im_out)
def orientedPaint(im, texture, N=7000, size=50, noise=0.3):
'''same as painterly but computes and uses the local orientation information to orient strokes.'''
importanceLow = np.ones_like(im)
outLow = io.constantIm( im.shape[0], im.shape[1], 0.0 )
thetas = computeAngles(im)
singleScaleOrientedPaint(im, outLow, thetas, importanceLow, texture, size, N, noise)
importanceHigh = helper.sharpnessMap(im)
singleScaleOrientedPaint(im, outLow, thetas, importanceHigh, texture, size/4, N, noise)
return outLow
def scaleNN(im, k):
h, w = im.shape[0:2]
out = imageIO.constantIm(h*k, w*k, 0.0)
for y,x in imIter(out):
try:
y_orig, x_orig = int(y/k), int(x/k)
out[y,x] = im[y_orig, x_orig]
except: #only for testing purposes
print y_orig, x_orig
return out
def basicDemosaic(raw, offsetGreen=0, offsetRedY=1, offsetRedX=1, offsetBlueY=0, offsetBlueX=0):
'''takes a raw image and a bunch of offsets. Returns an rgb image computed with our basic techniche.'''
height = raw.shape[0]
width = raw.shape[1]
out = io.constantIm(height, width, [0,0,0])
out[:,:,0] = basicRorB(raw, offsetRedY, offsetRedX)
out[:,:,1] = basicGreen(raw) #use default offsetGreen
out[:,:,2] = basicRorB(raw, offsetBlueY, offsetBlueX)
return out
def edgeBasedGreenDemosaic(raw, offsetGreen=0, offsetRedY=1, offsetRedX=1, offsetBlueY=0, offsetBlueX=0):
'''same as basicDemosaic except it uses the edge based technique to produce the green channel.'''
height = raw.shape[0]
width = raw.shape[1]
out = io.constantIm(height, width, [0,0,0])
out[:,:,0] = basicRorB(raw, offsetRedY, offsetRedX)
out[:,:,1] = edgeBasedGreen(raw) #use default offsetGreen
out[:,:,2] = basicRorB(raw, offsetBlueY, offsetBlueX)
return out
def stitch(im1, im2, listOfPairs):
H = computeHomography(listOfPairs) # pairs from 1->2
bbox = computeTransformedBBox(im1, H)
t_m = translate(bbox)
t_m_inv = np.linalg.inv(t_m)
h, w = bbox[1][0] - bbox[0][0], bbox[1][1] - bbox[0][1]
out = imageIO.constantIm(h, w, [0, 0, 0.0])
applyHomography(im1, out, t_m, bilinear=True)
comb = t_m+H
# applyhomography(im2, out, comb, bilinear=True)
return out
def improvedDemosaic(raw, offsetGreen=0, offsetRedY=1, offsetRedX=1, offsetBlueY=0, offsetBlueX=0):
'''Same as basicDemosaic but uses edgeBasedGreen and greenBasedRorB.'''
height = raw.shape[0]
width = raw.shape[1]
out = io.constantIm(height, width, [0,0,0])
green = edgeBasedGreen(raw) #use default offsetGreen
out[:,:,0] = greenBasedRorB(raw, green, offsetRedY, offsetRedX)
out[:,:,1] = green
out[:,:,2] = greenBasedRorB(raw, green, offsetBlueY, offsetBlueX)
return out
def epiSlice(LF, y):
'''Takes a light field. Returns the epipolar slice with constant v=(nv/2) and constant y (input argument).'''
nv = LF.shape[0]
nu = LF.shape[1]
ny = LF.shape[2]
nx = LF.shape[3]
v=(nv/2)
out = io.constantIm( nu, nx, 0.0)
for x in range(nx):
for u in range(nu):
out[u, x] = LF[v,u,y,x]
return out
def rotate(im, theta):
'''takes an image and an angle in radians as input
returns an image of the same size and rotated by theta
'''
out = io.constantIm(im.shape[0], im.shape[1], 0)
halfX = int(im.shape[1] / 2)
halfY = int(im.shape[0] / 2)
inverseRotateM = np.array([[math.cos(theta), -math.sin(theta)],
[math.sin(theta), math.cos(theta)]])
for y, x in imIter(out):
origImgPos = np.dot(inverseRotateM, np.array([x - halfX, y - halfY])) + np.array([halfX, halfY])
out[y, x] = interpolateLin(im, origImgPos[1], origImgPos[0])
return out
def split(raw):
'''splits one of Sergei's images into a 3-channel image with height that is floor(height_of_raw/3.0). Returns the 3-channel image.'''
width = raw.shape[1]
height = raw.shape[0]
crop = math.floor(height/3.0)
out = io.constantIm(crop, width, [0,0,0])
out[:,:,0] = raw[2*crop:3*crop,:]
out[:,:,1] = raw[crop:2*crop,:]
out[:,:,2] = raw[:crop,:]
return out
def apertureView(LF):
'''Takes a light field, returns 'out,' an image with nx*ny sub-pictures representing the value of each pixel in each of the nu*nv views.'''
nv = LF.shape[0]
nu = LF.shape[1]
ny = LF.shape[2]
nx = LF.shape[3]
out = io.constantIm( nv*ny, nu*nx, 0.0)
for y in range(ny):
for x in range(nx):
for v in range(nv):
for u in range(nu):
out[nv*y+v,nu*x+u] = LF[v,u,y,x]
return out
def warpBy1(im, segmentBefore, segmentAfter):
'''Takes an image, one before segment, and one after segment.
Returns an image that has been warped according to the two segments.
'''
height = im.shape[0]
width = im.shape[1]
out = io.constantIm(height, width, 0.0)
for y,x in imIter(out):
X = np.array([y,x], dtype=np.float64)
(u,v) = segmentAfter.uv(X)
X_prime = segmentBefore.uvtox(u,v)
pixel = interpolateLin(im, X_prime[0], X_prime[1])
out[y,x] = pixel
return out
def merge2images(im1, im2, H):
B = computeTransformedBBox(im1, np.linalg.inv(H))
B = bboxUnion(B, np.array([[0.0, 0.0], 1.0 * np.array(im2.shape[0:2])]))
T = translate(B[0, 0], B[0, 1])
out = imageIO.constantIm(B[1, 0] - B[0, 0], B[1, 1] - B[0, 1], 0)
ws1 = calcLinBlendWeights(im1)
ws2 = calcLinBlendWeights(im2)
H2 = np.dot(H, T)
applyDoubleHomography(im1, im2, ws1, ws2, out, T, H2, bilinear=False)
# applyHomography(im1, out, T)
# applyHomography(im2, out, np.dot(H, T))
return out
def compositeNImages(listOfImages, listOfH, listOfWeights=None, twoScale=False):
'''Computes the composite image. listOfH is of the form returned by computeNHomographies. Hint: You will need to
deal with bounding boxes and translations again in this function.'''
bbox = getOuterBBox( listOfImages, listOfH )
trans = translate(bbox)
height = bbox[1][0] - bbox[0][0] + 1
width = bbox[1][1] - bbox[0][1] + 1
out = io.constantIm( height, width, 0.0)
if (listOfWeights is not None):
weightSum = io.constantIm( height, width, 0.0)
if (twoScale):
sigmaG = 2.0
L_low = map(lowFreqIm, listOfImages, [sigmaG] * len(listOfImages))
L_high = map(highFreqIm, listOfImages, [sigmaG] * len(listOfImages))
weightSumL = io.constantIm( height, width, 0.0)
weightSumH = io.constantIm( height, width, 0.0)
outL = io.constantIm( height, width, 0.0)
outH = io.constantIm( height, width, 0.0)
for i in range(len(listOfH)):
H = np.dot( listOfH[i], trans)
if (listOfWeights is not None):
w = listOfWeights[i]
imW = listOfImages[i].copy()
imW[:,:,0] = w
imW[:,:,1] = w
imW[:,:,2] = w
if( twoScale ):
imL = L_low[i]
applyHomographyFastBlended(imL, outL, H, True, imW, weightSumL)
imH = L_high[i]
applyHomographyFastBlended(imH, outH, H, True, imW, weightSumH, True)
else:
im = listOfImages[i]
applyHomographyFastBlended(im, out, H, True, imW, weightSum)
else:
applyHomographyFast(listOfImages[i], out, H, True)
if (listOfWeights is not None):
if( twoScale ):
#io.imwrite(weightSumL, 'debugWeightsLow.png', 1.0)
#io.imwrite(weightSumH, 'debugWeightsHigh.png', 1.0)
out = outL + outH
return out
def scaleBiquadratic(im, k):
'''Takes an image and a scale factor. Returns an image scaled using bilinear interpolation.
'''
height = im.shape[0]
width = im.shape[1]
newHeight = height*k
newWidth = width*k
out = io.constantIm(newHeight, newWidth, 0.0)
for y,x in imIter(out):
x_float = x/float(newWidth) * (width)
y_float = y/float(newHeight) * (height)
pix = interpolateBiquad(im, y_float, x_float)
out[y,x] = pix
return out
def orientedPaint(im, texture, N=7000, size=50, noise=0.3):
'''same as painterly but computes and uses the local orientation information to orient strokes.'''
out = io.constantIm(im.shape[0], im.shape[1])
thetas = computeAngles(im)
# first pass
importance_first_pass = np.ones_like(im)
singleScaleOrientedPaint(im, out, thetas, importance_first_pass, texture,
size, N, noise)
# second pass
importance_second_pass = helper.sharpnessMap(im)
singleScaleOrientedPaint(im, out, thetas, importance_second_pass, texture,
size / 4, N, noise)
return out
def rotate(im, theta):
'''takes an image and an angle in radians as input
returns an image of the same size and rotated by theta
'''
h = im.shape[0]
w = im.shape[1]
center = (int(h/2), int(w/2))
out = io.constantIm(h, w, 0.0)
for y,x in imIter(out):
tx = ((x-center[1])*math.cos(theta) - (y-center[0])*math.sin(theta) + center[1])
ty = ((x-center[1])*math.sin(theta) + (y-center[0])*math.cos(theta) + center[0])
pix = interpolateLin(im, ty, tx, True)
out[y,x] = pix
return out