def image_fisher_featurize_sift(im_keys, out_matrix, bidx, gmm_sift,
pca_sift_mat, pca_sift_mu):
features = []
t = time.time()
for im_key in im_keys:
s3 = boto3.resource('s3')
client = boto3.client('s3')
bio = io.BytesIO(
client.get_object(Bucket="pictureweb", Key=im_key)["Body"].read())
im = scipy.misc.imread(bio, flatten=True)
im = im.astype('float32')
im /= 255.0
descs = sift.sift(im)
descs = (descs).dot(pca_sift_mat)
sift_features = fisher.fisher_vector_features(descs.astype('float32'),
*gmm_sift)
features.append(sift_features)
#sqrt normalization
signs = np.sign(features)
features = signs * np.sqrt(np.abs(features))
feature_norms = np.linalg.norm(features, axis=1)[:, np.newaxis]
features /= feature_norms
out_matrix.put_block(features, bidx, 0)
e = time.time()
return e - t
def init_photos_reconstruction(dirpath, str1, str2):
# 读入图像
image1 = cv2.imread(dirpath + str1)
image2 = cv2.imread(dirpath + str2)
# image1 = cv2.resize(image1, (1000, 750))
# image2 = cv2.resize(image2, (1000, 750))
# sift、surf + KDTree
sift_pts = sift.sift(image1, image2)
surf_pts = surf.surf(image1, image2)
pts = sift_pts + surf_pts
# pts使用RANSAC, 返回RANSAC后的匹配点、本征矩阵、R、t矩阵、匹配点以及相机内参
pts, essentialMat, R, t, src_pts, dst_pts, K = RANSAC.compute(pts)
# 透视变换并保存
# image3 = cv2.warpPerspective(image1, M, (1000, 720))
# cv2.imwrite("warpPerspective.tif", image3)
global projMatr, RList, tList, points, colors
points4D, color4D, Mat = sfm.triangulate_compute(projMatr, R, t, src_pts, dst_pts, K, image1)
projMatr = Mat
RList.append(projMatr[:, 0:3])
tList.append(projMatr[:, 3:4])
points.append(points4D)
colors.append(color4D)
# 水平拼接
# image = np.hstack((image1, image2))
# 存储点云信息
# save.save_data(image1, src_pts, R, t, points4D)
# 显示匹配线段
# sift.show_line(image, pts)
# cv2.waitKey()
# 显示点云
# os.system('show')
return dst_pts, points4D
def main():
print "opening image"
im = Image.open("test.pgm")
l = list(im.getdata())
l.insert(0,im.size)
#the image data must be a list of pixels in row order
#(that is, the first row, then the second row etc)
#with a (width,height) tuple as the first element
#the image must be grayscale (one channel)
print "running sift"
#this will calculate the keypoints and then the sift descriptors for each
desc = sift.sift(l,verbose=True)
print "done"
print len(desc), "descriptors found"
#print desc
output = open('results.txt','w')
for a in desc:
x = a[0]
y = a[1]
im.putpixel((x,y),int(255))
printDescriptor(a,output)
output.close()
im.show()
def generateSift(self, imgOrFilename): #if isinstance(imgOrFilename, str): #grayImg = cv2.imread(imgOrFilename, cv2.CV_LOAD_IMAGE_GRAYSCALE) return sift.sift().calculateSiftDescriptorFromFile(imgOrFilename, self.gridSpacing, self.patchSize, self.maxImSize, self.nrml_threshold) #grayImg = grayImg = cv2.cvtColor(imgOrFilename, cv2.COLOR_BGR2GRAY) #else: #grayImg = cv2.cvtColor(imgOrFilename, cv2.COLOR_BGR2GRAY) #return sift.sift().calculateSiftDescriptor(imgOrFilename, self.gridSpacing, self.patchSize, self.maxImSize, self.nrml_threshold)
def extract_features(image_paths):
Features = [] # 所有图像的所有sift descriptor的列表
for i, image_path in enumerate(image_paths):
img = cv.imread(image_path)
gray = cv.cvtColor(img, cv.COLOR_RGB2GRAY)
kp, img_des = sift.sift(image_path)
Features.append(img_des)
return Features
def image_fisher_featurize(im_key, gmm_sift, gmm_lcs, pca_sift, pca_lcs):
t = time.time()
s3 = boto3.resource('s3')
s3.Bucket("pictureweb").download_file(im_key, "/tmp/img.jpg")
im = scipy.misc.imread("/tmp/img.jpg", flatten=True)
descs = sift(im).dot(pca_sift.T)
sift_features = fisher.fisher_vector_features(descs.astype('float32'), *gmm_sift)
im = scipy.misc.imread("/tmp/img.jpg")
descs = lcs(im).reshape(-1, 96).dot(pca_lcs.T)
lcs_features = fisher.fisher_vector_features(descs.astype('float32'), *gmm_lcs)
out_features = np.hstack((sift_features, lcs_features)).T
e = time.time()
return out_features, e - t
def demo(imgfile0,imgfile1,rth0,rth1,mth):
# read image, convert to gray
#img = io.imread(sys.argv[1]);
img=io.imread("data/plane.bmp")
img0 = io.imread(imgfile0)
img1 = io.imread(imgfile1)
print("orig0 shape: "+str(img0.shape))
print("orig1 shape: "+str(img1.shape))
gry0 = rgb2gray(img0)
gry1 = rgb2gray(img1)
print("gray0 shape: "+str(gry0.shape))
print("gray1 shape: "+str(gry1.shape))
# get list of locations of interest points
feat0 = harris(gry0,gk(3,3,1),rth0)
feat1 = harris(gry1,gk(3,3,1),rth1)
print("number of features0: "+str(len(feat0)))
print("number of features1: "+str(len(feat1)))
# make sift descriptors
des0 = sift(gry0,feat0)
des1 = sift(gry1,feat1)
showFeatures(img0,feat0,des0)
showFeatures(img1,feat1,des1)
M = matchSIFT(des0,des1,40,mth)
img2 = drawMatches(img0,feat0,img1,feat1,M)
io.imshow( img2,vmin=0,vmax=255,cmap="gray")
io.show()
def main():
print "opening image"
im = Image.open("test.pgm")
l = list(im.getdata())
l.insert(0,im.size)
#the image data must be a list of pixels in row order
#(that is, the first row, then the second row etc)
#with a (width,height) tuple as the first element
#the image must be grayscale (one channel)
print "running sift"
#we do this to load in a file of keypoints so that sift
#doesnt have to find them (or we can guide it)
#takes the first 4 floats per line and puts in a tuple
input = open('keypoints.txt','r')
keys = []
for a in input:
keys.append(tuple(map(float,a.split()[0:4])))
input.close()
#keys must be a list of 4-tuples for x,y,sigma and theta of each keypoint
desc = sift.sift(l,keypoints=keys,verbose=True)
print "done"
print len(desc), "descriptors found"
#print desc
output = open('results.txt','w')
for a in desc:
x = a[0]
y = a[1]
im.putpixel((x,y),int(255))
printDescriptor(a,output)
output.close()
im.show()
def calculate_sifts(img_keys, out_matrix, block_idx, descs_per_img=16):
sifts = []
s3 = boto3.resource('s3')
import time
t = time.time()
np.random.seed(block_idx)
for im_key in img_keys:
s3.Bucket("pictureweb").download_file(im_key, "/tmp/img.jpg")
im = scipy.misc.imread("/tmp/img.jpg", flatten=True)
im = im.astype('float32')
im /= 255.0
descs = sift.sift(im)
assert (np.all(descs >= 0))
idxs = np.random.choice(descs.shape[0], descs_per_img)
descs = np.sqrt(descs)
sifts.append(descs[idxs, :])
out_matrix.put_block(np.vstack(sifts), block_idx, 0)
e = time.time()
return e - t
def starter():
image_id, model = input("Enter image ID and model name (LBP or SIFT): ").split()
constants_dict = constants.read_json()
read_path = constants_dict["READ_PATH"]
write_path = constants_dict["WRITE_PATH"]
files = os.listdir(read_path)
file = files[files.index("{}.jpg".format(image_id))]
img = cv2.imread(read_path + file)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
if model == "LBP":
print(lbp.lbp(gray))
else:
print(sift.sift(gray))
print("Done... ")
def photos_reconstruction(dirpath, str1, str2, pre_dst_pts, pre_points4D):
# 读入图像
image1 = cv2.imread(dirpath + str1)
image2 = cv2.imread(dirpath + str2)
# image1 = cv2.resize(image1, (1000, 750))
# image2 = cv2.resize(image2, (1000, 750))
# sift、surf + KDTree
sift_pts = sift.sift(image1, image2)
surf_pts = surf.surf(image1, image2)
pts = sift_pts + surf_pts
# pts使用RANSAC, 返回RANSAC后的匹配点、本征矩阵、R、t矩阵、匹配点以及相机内参
pts, essentialMat, R, t, src_pts, dst_pts, K = RANSAC.compute(pts)
src_pts_PnP, src_pts_PnP4D = kdtree.KDTree_match_pre(pre_dst_pts, pre_points4D, src_pts, dst_pts)
# print(src_pts_PnP, src_pts_PnP4D)
retval, rvec, t, inliers = cv2.solvePnPRansac(src_pts_PnP4D, src_pts_PnP, K, np.zeros(4), flags=cv2.SOLVEPNP_EPNP, reprojectionError=8.0, confidence=0.99)
R = cv2.Rodrigues(rvec)[0]
# 透视变换并保存
# image3 = cv2.warpPerspective(image1, M, (1000, 720))
# cv2.imwrite("warpPerspective.tif", image3)
global projMatr, RList, tList, points, colors
print(t)
# print(projMatr[:, 3:4])
points4D, color4D, Mat = sfm.triangulate_compute_pre(projMatr, R, t, src_pts, dst_pts, K, image1)
projMatr = Mat
RList.append(projMatr[:, 0:3])
tList.append(projMatr[:, 3:4])
points.append(points4D)
colors.append(color4D)
# 水平拼接
# image = np.hstack((image1, image2))
# 存储点云信息
# save.save_data(image1, src_pts, R, t, points4D)
# 显示匹配线段
# sift.show_line(image, pts)
# cv2.waitKey()
# 显示点云
# os.system('show')
return dst_pts, points4D
def store_feature_vectors(collection, read_path, write_path):
files = os.listdir(read_path)
for file in files:
print("Reading file: {}".format(file))
img = cv2.imread(read_path + file)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
lbp_results = lbp.lbp(gray)
sift_results = sift.sift(gray)
feature_vector = {
"name": file.replace(".jpg", ""),
"lbp": lbp_results.tolist(),
"sift": sift_results.tolist()
}
with open("{}{}_fd.json".format(write_path, file.replace(".jpg", "")),
"w") as fp:
json.dump(feature_vector, fp, indent=4, sort_keys=True)
feature_vector["lbp"] = Binary(pickle.dumps(lbp_results))
feature_vector["sift"] = Binary(pickle.dumps(sift_results))
collection.insert_one(feature_vector)
means = convert_keystone_csv_to_numpy("./gmmMeans_lcs.csv").T
covars = convert_keystone_csv_to_numpy("./gmmVars_lcs.csv").T
descs = pca_mat.dot(descs)
pca_keystone = convert_keystone_csv_to_numpy("./pca_keystone_lcs.txt")
gmm = (means, covars, weights)
fv_keystone = convert_keystone_csv_to_numpy("./fisher_keystone_lcs.txt")
fv_features = fisher.fisher_vector_features(
descs.astype('float32').T, *gmm)
print("AVG LCS DIFF ", np.mean(np.abs(fv_features - fv_keystone)))
print("MAX LCS DIFF ", np.max(np.abs(fv_features - fv_keystone)))
im = scipy.misc.imread("./ILSVRC2012_val_00000293.JPEG", flatten=True)
im /= 255.0
descs = sift.sift(im).T
descs_keystone = C_ARRAY(
convert_keystone_csv_to_numpy("./sift_imagenet.txt"))
pca_mat = convert_keystone_csv_to_numpy("./pcaMat_sift.csv")
weights = convert_keystone_csv_to_numpy("./gmmCoefs_sift.csv")
means = convert_keystone_csv_to_numpy("./gmmMeans_sift.csv").T
covars = convert_keystone_csv_to_numpy("./gmmVars_sift.csv").T
descs = pca_mat.dot(descs)
pca_keystone = convert_keystone_csv_to_numpy("./pca_keystone_sift.txt")
gmm = (means, covars, weights)
fv_keystone = convert_keystone_csv_to_numpy("./fisher_keystone_sift.txt")
fv_features = fisher.fisher_vector_features(
descs.astype('float32').T, *gmm)
import sift
if __name__ == '__main__':
s = sift.sift()
finre, scal = s.match("./img/1.png", "./img/2.png")
for i in finre:
print(i[0])
print(i[1])
print(scal)
def image_fisher_featurize_sift_lcs(im_keys, out_matrix, bidx, gmm_sift,
pca_sift_mat, gmm_lcs, pca_lcs_mat):
all_sift_features = []
all_lcs_features = []
t = time.time()
for im_key in im_keys:
s3 = boto3.resource('s3')
client = boto3.client('s3')
bio = io.BytesIO(
client.get_object(Bucket="pictureweb", Key=im_key)["Body"].read())
im = scipy.misc.imread(bio, flatten=True)
im = im.astype('float32')
im /= 255.0
sift_descs = sift.sift(im)
assert (np.all(sift_descs >= 0))
sift_descs = np.sqrt(sift_descs)
sift_descs = (sift_descs).dot(pca_sift_mat)
sift_features = fisher.fisher_vector_features(
sift_descs.astype('float32'), *gmm_sift)
sift_features /= np.linalg.norm(sift_features)
bio = io.BytesIO(
client.get_object(Bucket="pictureweb", Key=im_key)["Body"].read())
im = scipy.misc.imread(bio)
lcs_descs = lcs.lcs(im).reshape(-1, LCS_DESC_LENGTH)
try:
assert (np.any(np.isnan(lcs_descs)) == False)
except:
raise Exception("RAISING LCS Error pre pca in {0}".format(im_key))
lcs_descs = (lcs_descs).dot(pca_lcs_mat)
try:
assert (np.any(np.isnan(lcs_descs)) == False)
except:
raise Exception("RAISING LCS Error post pca in {0}".format(im_key))
lcs_features = fisher.fisher_vector_features(
lcs_descs.astype('float32'), *gmm_lcs)
lcs_features /= np.linalg.norm(lcs_features)
try:
assert (np.any(np.isnan(lcs_features)) == False)
except:
raise Exception(
"RAISING LCS Fisher Vector Error in {0}".format(im_key))
all_sift_features.append(sift_features)
all_lcs_features.append(lcs_features)
all_sift_features = np.array(all_sift_features)
all_lcs_features = np.array(all_lcs_features)
assert (np.any(np.isnan(all_sift_features)) == False)
#sqrt normalization
signs = np.sign(all_sift_features)
all_sift_features = signs * np.sqrt(np.abs(all_sift_features))
feature_norms = np.linalg.norm(all_sift_features, axis=1)[:, np.newaxis]
assert (np.any(np.isnan(feature_norms)) == False)
all_sift_features /= feature_norms
assert (np.any(np.isnan(all_sift_features)) == False)
assert (np.any(np.isnan(all_lcs_features)) == False)
signs = np.sign(all_lcs_features)
all_lcs_features = signs * np.sqrt(np.abs(all_lcs_features))
feature_norms = np.linalg.norm(all_lcs_features, axis=1)[:, np.newaxis]
assert (np.any(np.isnan(feature_norms)) == False)
all_lcs_features /= feature_norms
assert (np.any(np.isnan(all_lcs_features)) == False)
features = np.hstack((all_sift_features, all_lcs_features))
out_matrix.put_block(features, bidx, 0)
e = time.time()
return e - t, t, e
def printDescriptor(desc,file):
map(lambda a: file.write(str(a)+' '),desc)
file.write('\n')
if 0:
a = np.zeros((128,128),dtype=np.uint8)
a[40:70,20:50] = 255
im = Image.fromarray(a,mode='L')
else:
im = Image.open(input_image)
a = np.asarray(im)
a = np.flipud(a)
l = list(im.getdata())
l.insert(0,im.size)
desc = sift.sift(l,verbose=True)
print len(desc), "descriptors found"
output = open(output_descriptors,'w')
pl.close()
pl.imshow(np.asarray(im),cmap="gray", origin='upper')
pl.axis('off')
r = np.linspace(0,2*math.pi,100)
for a in desc:
x,y,scale,ang = a[:4]
im.putpixel((x,y),int(255))
pl.plot(x+scale/2*np.cos(r),y+scale/2*np.sin(r),'r')
pl.plot((x,x+scale/2*math.cos(ang)),(y,y+scale/2*math.sin(ang)),'g')