In this post, I would like to introduce my implementation of "object co-localization" proposed in
"Co-localization in Real World Images" [Tang+, CVPR2014].
In particular, I used this co-localization for face recognition. The method aims to find specific people who commonly appear in an image set. Object proposals are generated via face detection. Prior is calculated as the probability of skin pixels.
This implementation requires IBM CPlex to solve a binary quadratic programming problem. As suggested in the original paper, you could rely on a continuous QP instead by modifying __disc_clustering function.
colocalize.py
'''
Object co-localization for face recognition
Jan 30, 2015
@jellied_unagi
'''
from pycpx import CPlexModel
import cv2
from skimage.io import ImageCollection
from skimage.feature import local_binary_pattern
from skimage.color import rgb2gray
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import AffinityPropagation
import numpy as np
import matplotlib.pyplot as plt
class colocalize():
'''
This class implements "Co-localization in Real World Images" [Tang+, CVPR2014] for localizing people
who commonly appear in an image set.
- Object proposals are generated via face detection.
- Prior is calculated as the probability of skin pixels.
- Optimization is done based on IBM CPLex withouth convex relaxation though the original implementation
relies on continuous QP instead of BQP.
'''
def __init__(self, fsize=100, mn=3, th=0.3, lbp_rad=8, lbp_max=66, method='default', kappa=1e-3, alpha=1e-3):
# Face detector for object proposals
self.detector = [cv2.CascadeClassifier(x)
for x in ['haarcascade_frontalface_default.xml',
'haarcascade_frontalface_alt.xml',
'haarcascade_frontalface_alt2.xml',
'haarcascade_frontalface_alt_tree.xml',
'haarcascade_profileface.xml']]
self.fsize = fsize # minimum face size
self.mn = mn # minimum support for detecting faces (larger for more precise detection)
self.th = th # threshold for non-maximum supporession
# Parameters for local binary pattern histograms
self.lbp_rad = lbp_rad
self.lbp_np = 8 * lbp_rad
self.lbp_max = lbp_max
self.method = method
self.ss = StandardScaler()
# Parameters for co-localization
self.kappa = kappa # ridge parameter
self.alpha = alpha # importance of prior
# Data holder
self.image_list = []
self.face_list = []
self.feat_list = []
self.prior_list = []
self.result_list = []
def register(self, image):
'''
Registering an image to the list
'''
print 'Registering image...',
im_gray = rgb2gray(image)
face = self.__detect_face(image, size=self.fsize, mn=self.mn, th=self.th)
if(len(face) == 0):
print 'Cound not find faces'
return 0
skin = self.__detect_skin(image)
lbp = local_binary_pattern(im_gray, self.lbp_np, self.lbp_rad, self.method)
feat = []
prior = []
for f in face:
feat.append(np.histogram(lbp[f[1]:f[3], f[0]:f[2]].ravel(),
self.lbp_max, normed=True)[0])
prior.append(np.mean(skin[f[1]:f[3], f[0]:f[2]].ravel() / 255.))
self.image_list.append(image)
self.face_list.append(face)
self.feat_list.append(np.vstack(feat))
self.prior_list.append(np.vstack(prior))
print 'done.'
def localize(self):
'''
Performing co-localization on the registered images
'''
# Scaling features
self.ss.fit(np.vstack(self.feat_list))
feat_list = [self.ss.transform(x.copy()) for x in self.feat_list]
print 'Solving BQP ...',
idx = self.__disc_clustering(feat_list, self.prior_list)
self.result_list = [x[y] for (x, y) in zip(self.face_list, idx)]
print 'done.'
self.show_results()
def show_results(self, is_all=True):
'''
Visualizing results
'''
plt.figure(figsize=(16, 16))
n_images = len(self.image_list)
for i in range(n_images):
if(is_all):
plt.subplot(np.sqrt(n_images) + 1, np.sqrt(n_images) + 1, i + 1)
else:
plt.show()
plt.figure(figsize=(16, 16))
img = self.image_list[i]
face = self.result_list[i]
plt.imshow(img)
for f in self.face_list[i]:
plt.plot([f[0], f[0], f[2], f[2], f[0]],
[f[1], f[3], f[3], f[1], f[1]], 'b', lw=6)
plt.plot([face[0], face[0], face[2], face[2], face[0]],
[face[1], face[3], face[3], face[1], face[1]], 'r', lw=6)
plt.axis('off')
def __detect_face(self, image, size=80, mn=1, th=0.3):
'''
Running the VJ face detector implemented in OpenCV
'''
face = [x.detectMultiScale(image, scaleFactor=1.1, minNeighbors=mn,
minSize=(size, size), flags=cv2.cv.CV_HAAR_SCALE_IMAGE) for x in self.detector]
if (np.sum([len(x) for x in face]) == 0):
if(size==0):
print 'Could not find faces'
return []
else:
size = np.max((size - 10, 10))
mn = np.max((mn - 1, 1))
print 'searching face...(size %d)' % size
return self.__detect_face(image, size=size, mn=mn, th=th)
face = np.vstack([x for x in face if len(x) > 0])
face[:, 2:] += face[:, :2]
face = self.__nms(face, th=th)
return face
def __nms(self, face, th=.3):
'''
non-maximum suppression of detected faces
'''
n_faces = len(face)
fzero = np.zeros((np.max(face[:, 3]), np.max(face[:, 2])))
fmat = []
fsum = []
for i in range(n_faces):
tmp = fzero.copy()
tmp[face[i, 1]:face[i, 3], face[i, 0]:face[i, 2]] = 1
fmat.append(tmp.ravel().astype('bool'))
fsum.append(np.sum(tmp))
rem = np.ones(n_faces)
for i in range(n_faces):
for j in range(n_faces):
if i != j:
fand = np.sum(fmat[i] & fmat[j])
if((fsum[i] < fsum[j]) & ((fand * 1. / fsum[i]) > th)):
rem[i] = 0
return face[rem == 1, :]
def __disc_clustering(self, feat_list, prior_list):
'''
Performing discriminative clustering via BQP
'''
X = np.matrix(np.vstack(feat_list))
nb = X.shape[0] * 1.
I_nb = np.matrix(np.eye(X.shape[0]))
I1_nb = np.matrix(np.ones(X.shape[0])).T
I = np.matrix(np.eye(X.shape[1]))
cpmat = I_nb - I1_nb * I1_nb.T / nb
A = cpmat * (I_nb - X * np.linalg.inv(X.T * cpmat * X + nb * self.kappa * I) * X.T) * cpmat / nb
P = np.matrix(np.vstack(prior_list))
print np.max(A), np.max(P)
n_cand = np.array([len(x) for x in feat_list])
cand_idx = np.hstack((0, np.cumsum(n_cand)))
B = np.matrix(np.zeros((len(n_cand), A.shape[0])))
for i in range(len(cand_idx) - 1):
B[i, cand_idx[i]:cand_idx[i + 1]] = 1
m = CPlexModel()
U = m.new((A.shape[0]), vtype = bool)
b = np.ones(len(n_cand))
m.constrain(B * U == b)
m.minimize(U.T * A * U - self.alpha * P.T * U)
idx = np.argwhere(m[U]==1).flatten() - cand_idx[:-1]
return idx
def __detect_skin(self, image):
'''
Calculating prior
'''
lower = np.array([0, 48, 80], dtype = "uint8")
upper = np.array([20, 255, 255], dtype = "uint8")
hsv = cv2.cvtColor(image, cv2.cv.CV_RGB2HSV)
skinMask = cv2.inRange(hsv, lower, upper)
return skinMask