def PairDiscovery(img1Name,
img1Dir,
img2Name,
img2Dir,
model,
transform,
tolerance,
margin,
scaleImgRef,
scaleList,
eta=1e-7,
featLayer='conv4',
scoreType='ALL',
RefFeat={},
flip=False):
if scoreType == 'Identity':
ScorePos = outils.ScorePosIdentity
elif scoreType == 'Hough':
ScorePos = outils.ScorePosHough
elif scoreType == 'Affine':
ScorePos = outils.ScorePosAffine
strideNet = 16
minNet = 15
featChannel = 256 if featLayer == 'conv4' else 512
img1Path = os.path.join(img1Dir, img1Name)
I1 = Image.open(img1Path).convert('RGB')
feat1, pilImg1W, pilImg1H, feat1W, feat1H, list1W, list1H, img1Bbox = outils.FeatImgRef(
I1, scaleImgRef, minNet, strideNet, margin, transform, model,
featChannel, eta)
toleranceRef = tolerance / scaleImgRef
match1, match2, similarity, gridSize = outils.MatchPair(
minNet, strideNet, model, transform, scaleList, feat1, feat1W, feat1H,
img2Dir, img2Name, list1W, list1H, featChannel, tolerance, margin, eta,
RefFeat, flip)
if len(match2) < 3:
return 0., {}
bestScore, inlier = ScorePos(match1, match2, similarity, gridSize,
toleranceRef)
return bestScore, inlier
def PairDiscovery(featQuery, img2Path, model, transform, tolerance, minFeatCC,
margin, scaleImgRef, scaleList, houghInitial, nbSamplePoint,
nbIter, saveQuality, computeSaliencyCoef, out1, out2):
strideNet = 16
minNet = 15
featChannel = 256
feat1, I1, pilImg1W, pilImg1H, feat1W, feat1H, list1W, list1H, img1Bbox = featQuery
vote = outils.VoteMatrix(tolerance)
toleranceRef = tolerance / scaleImgRef
I2 = Image.open(img2Path).convert('RGB')
pilImg2W, pilImg2H = I2.size
match1, match2, similarity, matchSetT = outils.MatchPair(
minNet, strideNet, model, transform, scaleList, feat1, feat1W, feat1H,
I2, list1W, list1H, featChannel, tolerance, vote)
matchSetT = matchSetT if houghInitial else range(len(match1))
if len(matchSetT) < nbSamplePoint:
if out1:
SkipIteration(I1, I2, saveQuality, out1, out2)
return 0.
bestParams, bestScore, inlier = outils.RANSAC(nbIter, match1, match2,
matchSetT, similarity,
toleranceRef, nbSamplePoint)
if len(bestParams) == 0:
if out1:
SkipIteration(I1, I2, saveQuality, index)
return 0.
feat2W, feat2H = outils.FeatSizeImgTarget(bestParams, feat1W, feat1H)
if feat2W == 0 or feat2H == 0 or feat2W >= 1000 or feat2H >= 1000:
if out1:
SkipIteration(I1, I2, saveQuality, out1, out2)
return 0.
match1, match2, score = outils.BackwardVerification(
feat2W, feat2H, feat1W, feat1H, inlier)
finalMask1 = np.ones(
(img1Bbox[3] - img1Bbox[1], img1Bbox[2] - img1Bbox[0]),
dtype=np.uint8) * 100
finalMask2 = np.ones((pilImg2H, pilImg2W), dtype=np.uint8) * 100
mask2 = np.zeros((feat2W, feat2H))
mask1 = np.zeros((feat1W, feat1H))
match1, match2, score = outils.KeepOnlyLargeCC(match1, match2, mask1,
mask2, minFeatCC, score)
if len(match1) == 0:
if out1:
SkipIteration(I1, I2, saveQuality, out1, out2)
return 0.
_, score = outils.GetCC(match1, match2, mask1, mask2, score)
sumScore = np.sum(score)
finalScore = sumScore / float(feat1.size()[1])
if out1:
match1, match2 = np.array(outils.ExtendRemove(match1)), np.array(
outils.ExtendRemove(match2))
mask1[match1[:, 0], match1[:, 1]] = 1
mask2[match2[:, 0], match2[:, 1]] = 1
mask1 = imresize(mask1,
(finalMask1.shape[0], finalMask1.shape[1])) / 128 > 0
mask2 = imresize(mask2,
(finalMask2.shape[0], finalMask2.shape[1])) / 128 > 0
finalMask1[mask1] = 255
finalMask2[mask2] = 255
I1RGBA = cv2.cvtColor(np.array(I1), cv2.COLOR_RGBA2BGRA)
I2RGBA = cv2.cvtColor(np.array(I2), cv2.COLOR_RGBA2BGRA)
mask1Index = finalMask1 > 0
mask2Index = finalMask2 > 0
mask1 = np.ones((I1RGBA.shape[0], I1RGBA.shape[1])) * 100
mask1[img1Bbox[1]:img1Bbox[3], img1Bbox[0]:img1Bbox[2]] = finalMask1
mask2 = finalMask2
mask1, mask2 = mask1.astype(np.uint8), mask2.astype(np.uint8)
I1RGBA[:, :, 3] = mask1
I2RGBA[:, :, 3] = mask2
ratio1 = max(
max(I1RGBA.shape[0], I1RGBA.shape[1]) / float(saveQuality), 1)
I1RGBA = imresize(
I1RGBA,
(int(I1RGBA.shape[0] / ratio1), int(I1RGBA.shape[1] / ratio1)))
ratio2 = max(I2RGBA.shape[0], I2RGBA.shape[1]) / float(saveQuality)
I2RGBA = imresize(
I2RGBA,
(int(I2RGBA.shape[0] / ratio2), int(I2RGBA.shape[1] / ratio2)))
I1RGBA[:, :, 3] = BlurMask(I1RGBA[:, :, 3]).astype(np.uint8)
I2RGBA[:, :, 3] = BlurMask(I2RGBA[:, :, 3]).astype(np.uint8)
cv2.imwrite(out1, I1RGBA)
cv2.imwrite(out2, I2RGBA)
return finalScore