def preprocessing(self, imgs, msks, sz=(48, 48), augx=0):
print len(imgs)
if augx > 0:
print "augmenting train data ..."
# augx = 2xnSample+1
n_sample = np.int(augx / 2.0) - 1
imH, imW = imgs[0].shape[0:2]
borderH = np.int(imH * 0.2)
borderW = np.int(imW * 0.2)
w = imW - borderW
h = imH - borderH
x1s = np.random.randint(0, borderW, n_sample)
y1s = np.random.randint(0, borderH, n_sample)
imgs_crop = imgs
msks_crop = msks
for img, msk in zip(imgs, msks):
imgs_crop += [imcrop(img, [x1, y1, w, h]) for x1, y1 in zip(x1s, y1s)]
msks_crop += [imcrop(msk, [x1, y1, w, h]) for x1, y1 in zip(x1s, y1s)]
print len(imgs_crop)
imgs_flip = [pl.fliplr(im) for im in imgs_crop]
msks_flip = [pl.fliplr(im) for im in msks_crop]
imgs = imgs_crop + imgs_flip
msks = msks_crop + msks_flip
print len(imgs)
imgs_rs = [imresize(im, sz, interp="bicubic") for im in imgs]
imgs_norm = [imnormalize(im) for im in imgs_rs]
msks_norm = [imresize(im, sz, interp="bicubic") for im in msks]
imgs_final, msks_final = self.convert_data(imgs_norm, msks_norm)
print len(imgs_final)
return imgs_final, msks_final
def _sampling_patch(self, spidx, augx=0):
""" sampling local patches for predicting salience score """
# spf = lambda items, indices: map(lambda i: items[i], indices)
imgs = [imnormalize(im) for im in self.imgs[spidx]] # spf(self.imgs, spidx)
segmsks = self.segmsks[spidx]
salmsks = self.salmsks[spidx]
features = self.feats[spidx]
h, w = self.segmsks[0].shape
# l = np.int(self.patL/2.)
# sz = (self.patL, self.patL)
l = self.patL
r = np.int(l / 2)
imgids = [] # image indices
ctrids = [] # center grid indices
patches = [] # cropped patches
salscores = [] # corresponding salience scores
patfeats = [] # colorsift feature of cropped patch
for idx, img, seg, sal, feat in zip(spidx, imgs, segmsks, salmsks, features):
print idx
ids1 = []
ids2 = []
for i in range(len(self.xx)):
x = self.xx[i]
y = self.yy[i]
# if x-r>=0 and x+r<w and y-r>=0 and y+r<h and seg[y, x] > 0 :
if seg[y, x] > 0:
ids1.append(idx)
ids2.append(i)
# crop patches centering on these points
imgids += ids1
ctrids += ids2
patches += [imcrop(img, [self.xx[k] - r, self.yy[k] - r, l, l]) for k in ids2]
patfeats += [feat[k] for k in ids2]
salscores += [np.mean(imcrop(sal, [self.xx[k] - r, self.yy[k] - r, l, l])) for k in ids2]
# convert image to data format (normalize & roll axis) that is appropriate for training usage
imgids = np.asarray(imgids)
ctrids = np.asarray(ctrids)
patches = self._convert_data(patches)
patfeats = np.asarray(patfeats)
salscores = np.asarray(salscores)
os.system("free -hm")
return imgids, ctrids, patches, patfeats, salscores
def _convert_data(self, imgs):
""" this contains operations on images that make image become input data (cannot be shown) """
# convert from RGB to CIE-LAB color space
print "RGB to LAB conversion"
imgs = [imnormalize(im) for im in imgs]
# roll axis to put channel axis to the first
print "Roll channel axis"
imgs = [im.transpose((2, 0, 1)) for im in imgs]
# flatten images
print "Flatten images"
imgs = np.asarray(imgs)
imgs = imflatten(imgs)
# normalize to be zero mean and unit std
print "Zero-mean & unit std normalization"
imgs = normalize01(imgs)
return imgs
