def make_masks(image):
#hog_masks = HOG.make_feature_mask(image)
#hog_masks_shape = hog_masks.shape
#print(hog_masks.shape)
#hog_masks = hog_masks.reshape((hog_masks.shape[0] * hog_masks.shape[1], hog_masks.shape[2]))
#hog_masks_min = np.min(hog_masks, 0)
#hog_masks_max = np.max(hog_masks, 0)
#print('HoG Min[%s] Max[%s]' % (np.min(hog_masks_min), np.max(hog_masks_max)))
#print(hog_masks.shape)
#hog_masks = np.array(255 * np.clip(hog_masks, 0, 1), dtype=np.uint8)
#hog_masks,
#LBP.make_feature_mask(image_gray,
#pool_radius=3),
image = imfeat.convert_image(image, [{
'type': 'numpy',
'mode': 'bgr',
'dtype': 'uint8'
}])
image_gray = imfeat.convert_image(image, [{
'type': 'numpy',
'mode': 'gray',
'dtype': 'uint8'
}])
image_gradient = np.array(
GRAD.make_feature_mask(np.array(image_gray, dtype=np.float32) / 255.) *
255,
dtype=np.uint8)
image_lab = imfeat.convert_image(image, [{
'type': 'numpy',
'mode': 'lab',
'dtype': 'uint8'
}])
return np.ascontiguousarray(np.dstack([image, image_lab, image_gradient]),
dtype=np.uint8)
def test_name(self):
type_channel_modes = [('opencv', 'bgr', 'uint8'), ('pil', 'rgb', 'uint8'), ('numpy', 'bgr', 'uint8'),
('opencv', 'bgr', 'float32'), ('numpy', 'bgr', 'float32')]
to_np8 = lambda x: imfeat.convert_image(x, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'bgr'})
to_np32 = lambda x: imfeat.convert_image(x, {'type': 'numpy', 'dtype': 'float32', 'mode': 'bgr'})
for fn in ['lena.jpg', 'lena.pgm', 'lena.ppm']:
image_np8 = None
image_np32 = None
for i in load_images(fn):
if image_np8 is None:
image_np8 = to_np8(i)
image_np32 = to_np32(i)
np.testing.assert_equal(image_np8, np.array(image_np32 * 255, dtype=np.uint8))
for t, c, m in type_channel_modes:
cur_img = imfeat.convert_image(i, {'type': t, 'dtype': m, 'mode': c})
if t == 'opencv':
self.assertTrue(isinstance(cur_img, cv.iplimage))
elif t == 'pil':
self.assertTrue(Image.isImageType(cur_img))
else:
self.assertTrue(isinstance(cur_img, np.ndarray))
if m == 'uint8':
np.testing.assert_equal(image_np8, to_np8(cur_img))
else:
np.testing.assert_equal(image_np32, to_np32(cur_img))
def test_name(self):
type_channel_modes = [('opencv', 'bgr', 'uint8'), ('pil', 'rgb', 'uint8'), ('numpy', 'bgr', 'uint8'),
('opencv', 'bgr', 'float32'), ('numpy', 'bgr', 'float32')]
to_np8 = lambda x: imfeat.convert_image(x, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'bgr'})
to_np32 = lambda x: imfeat.convert_image(x, {'type': 'numpy', 'dtype': 'float32', 'mode': 'bgr'})
for fn in ['lena.jpg', 'lena.pgm', 'lena.ppm']:
image_np8 = None
image_np32 = None
for i in load_images(fn):
if image_np8 is None:
image_np8 = to_np8(i)
image_np32 = to_np32(i)
np.testing.assert_equal(image_np8, np.array(image_np32 * 255, dtype=np.uint8))
for t, c, m in type_channel_modes:
cur_img = imfeat.convert_image(i, {'type': t, 'dtype': m, 'mode': c})
if t == 'opencv':
self.assertTrue(isinstance(cur_img, cv.iplimage))
elif t == 'pil':
self.assertTrue(Image.isImageType(cur_img))
else:
self.assertTrue(isinstance(cur_img, np.ndarray))
if m == 'uint8':
np.testing.assert_equal(image_np8, to_np8(cur_img))
else:
np.testing.assert_equal(image_np32, to_np32(cur_img))
def map(self, image_hash, image_data):
"""
Args:
image_hash: Unique image string
image_data: Binary image data
Yields:
A tuple in the form of (classifier_name, label_value)
classifier_name: String representing the classifier
label_value: (label, feature) where label is an int
"""
try:
image = Image.open(StringIO.StringIO(image_data))
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
bgen = imfeat.BlockGenerator(image,
imfeat.CoordGeneratorRectRotate,
output_size=(self._image_height,
self._image_width),
step_delta=(self._image_height / 2,
self._image_width / 2),
angle_steps=1)
for num, (image_out, sim) in enumerate(bgen):
feature = np.asfarray(imfeat.compute(self._feat, image_out)[0])
pred = dict((classifier_name, classifier.predict(feature))
for classifier_name, classifier in self._classifiers)
if any(x for x in pred.values() if x[0][0] *
x[0][1] > 0): # At least 1 class needs to be > 0
image_out_fp = StringIO.StringIO()
imfeat.convert_image(image_out,
['RGB']).save(image_out_fp, 'JPEG')
image_out_fp.seek(0)
yield (image_hash, sim), (pred, image_out_fp.read())
def map(self, image_hash, image_data):
"""
Args:
image_hash: Unique image string
image_data: Binary image data
Yields:
A tuple in the form of (classifier_name, label_value)
classifier_name: String representing the classifier
label_value: (label, feature) where label is an int
"""
try:
image = Image.open(StringIO.StringIO(image_data))
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
bgen = imfeat.BlockGenerator(image, imfeat.CoordGeneratorRectRotate,
output_size=(self._image_height, self._image_width),
step_delta=(self._image_height / 2, self._image_width / 2), angle_steps=1)
for num, (image_out, sim) in enumerate(bgen):
feature = np.asfarray(imfeat.compute(self._feat, image_out)[0])
pred = dict((classifier_name, classifier.predict(feature))
for classifier_name, classifier in self._classifiers)
if any(x for x in pred.values() if x[0][0] * x[0][1] > 0): # At least 1 class needs to be > 0
image_out_fp = StringIO.StringIO()
imfeat.convert_image(image_out, ['RGB']).save(image_out_fp, 'JPEG')
image_out_fp.seek(0)
yield (image_hash, sim), (pred, image_out_fp.read())
def _test_convert(self, rgb_func, gray_func, is_rgb):
for image_name in image_names:
try:
image = rgb_func(image_name)
image_gray = gray_func(image_name)
except IOError:
# OpenCV doesn't like reading in animated gifs,
# or pallete gifs
continue
for mode in modes:
try:
# Explicitly ignore the case of converting Gray to RGB
if not is_rgb and not (mode == "L" or "gray" in mode):
continue
# Use either the gray or the RGB version as input based
# on the parameter
if is_rgb:
actual = imfeat.convert_image(image, [mode])
else:
actual = imfeat.convert_image(image_gray, [mode])
# Compare against the gray or RGB input depending on
# the output mode
if mode == "L" or "gray" in mode:
self._assert_close_rgb(actual, image_gray)
else:
self._assert_close_rgb(actual, image)
except:
print "Failed case: (convert to %s) %s [input %s]" % (mode, image_name, "rgb" if is_rgb else "gray")
raise
def save_display_images(path_hdfs, path_local, min_count,
max_count, key_to_path=None):
"""
Saves the first max_count images obtained by calling
hadoopy.readtb(path_hdfs). Each item in the sequence is assumed
to be of the form (key, (imagedata, boxes)). The boxes are
drawn on each image before it is saved to the local path.
If key_to_path is provided, which maps a key to a path, the image
corresponding to that key will be saved in key_to_path[key].
"""
if key_to_path == None:
key_to_path = {}
count = 0
for k, (i, bs) in hadoopy.readtb(path_hdfs):
if count >= min_count:
if k in key_to_path:
path = key_to_path[k]
else:
path = path_local
filename = '%s/%s.jpg' % (path, k)
im = imfeat.convert_image(Image.open(StringIO.StringIO(i)),
[('opencv', 'bgr', 8)])
print(k)
for b in bs:
cv.Rectangle(im, (b[0], b[1]), (b[2], b[3]),
cv.CV_RGB(255, 0, 0), 3)
cv.SaveImage(filename, im)
# update count and break loop if necessary
# TODO(Vlad): can we slice notation on a list of generators?
count += 1
if count > max_count:
break
def test_transparent_png(self):
"""Transparent PNGs read in by PIL (mode RGBA) aren't supported
"""
image = Image.open("test_images/transparent.png")
assert image.mode == "RGBA"
mode = ("opencv", "bgr", cv.IPL_DEPTH_8U)
actual = imfeat.convert_image(image, [mode])
self._assert_close_rgb(actual, image)
def test_save_lena(self):
feat = imfeat.GradientHistogram()
out = feat.make_feature_mask(imfeat.convert_image(cv2.imread('test_images/lena.jpg'), feat.MODES))
try:
os.makedirs('out')
except OSError:
pass
print('m[%s]M[%s]' % (np.min(out), np.max(out)))
cv2.imwrite('out/lena_gradient.jpg', np.array(out * 255, dtype=np.uint8))
def imread(self, fn):
out = cv2.imread(fn)
if out is not None:
return out
return imfeat.convert_image(Image.open(fn), {
'type': 'numpy',
'dtype': 'uint8',
'mode': 'bgr'
})
def make_masks(image):
#hog_masks = HOG.make_feature_mask(image)
#hog_masks_shape = hog_masks.shape
#print(hog_masks.shape)
#hog_masks = hog_masks.reshape((hog_masks.shape[0] * hog_masks.shape[1], hog_masks.shape[2]))
#hog_masks_min = np.min(hog_masks, 0)
#hog_masks_max = np.max(hog_masks, 0)
#print('HoG Min[%s] Max[%s]' % (np.min(hog_masks_min), np.max(hog_masks_max)))
#print(hog_masks.shape)
#hog_masks = np.array(255 * np.clip(hog_masks, 0, 1), dtype=np.uint8)
#hog_masks,
#LBP.make_feature_mask(image_gray,
#pool_radius=3),
image = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'bgr', 'dtype': 'uint8'}])
image_gray = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'gray', 'dtype': 'uint8'}])
image_gradient = np.array(GRAD.make_feature_mask(np.array(image_gray, dtype=np.float32) / 255.) * 255, dtype=np.uint8)
image_lab = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'lab', 'dtype': 'uint8'}])
return np.ascontiguousarray(np.dstack([image, image_lab, image_gradient]), dtype=np.uint8)
def test_tb(path):
"""
This function tests the sequence file at 'path' (on hdfs) by
reading the images from it.
"""
# test that we can read each file using _load_cv_image
for (key, val) in hadoopy.readtb(path):
print(key)
i = imfeat.convert_image(Image.open(StringIO.StringIO(val)),
[('opencv', 'gray', 8)])
def test_resize(self):
to_np8 = lambda x: imfeat.convert_image(x, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'bgr'})
n = 0
for fn in ['lena.jpg', 'lena.pgm', 'lena.ppm']:
for i in load_images(fn):
for h, w in [(50, 50), (100, 50), (50, 100), (1000, 100), (100, 1000)]:
out_arr = to_np8(imfeat.resize_image(i, h, w))
self.assertEqual(out_arr.shape, (h, w, 3))
#cv2.imwrite('resize-out-%.3d.jpg' % n, out_arr)
n += 1
def test_resize(self):
to_np8 = lambda x: imfeat.convert_image(x, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'bgr'})
n = 0
for fn in ['lena.jpg', 'lena.pgm', 'lena.ppm']:
for i in load_images(fn):
for h, w in [(50, 50), (100, 50), (50, 100), (1000, 100), (100, 1000)]:
out_arr = to_np8(imfeat.resize_image(i, h, w))
self.assertEqual(out_arr.shape, (h, w, 3))
#cv2.imwrite('resize-out-%.3d.jpg' % n, out_arr)
n += 1
def test_tb(path):
"""
This function tests the sequence file at 'path' (on hdfs) by
reading the images from it.
"""
# test that we can read each file using _load_cv_image
for (key, val) in hadoopy.readtb(path):
print(key)
i = imfeat.convert_image(Image.open(StringIO.StringIO(val)),
[('opencv', 'gray', 8)])
def test_save_lena(self):
feat = imfeat.LBP()
for x in range(10):
out = feat.make_feature_mask(imfeat.convert_image(cv2.imread('test_images/lena.jpg'), feat.MODES), pool_radius=x)
print(out.shape)
out = out.reshape(out.shape[:2])
try:
os.makedirs('out')
except OSError:
pass
cv2.imwrite('out/lena_lbp-%d.jpg' % x, out)
def test_save_lena(self):
feat = imfeat.GradientHistogram()
out = feat.make_feature_mask(
imfeat.convert_image(cv2.imread('test_images/lena.jpg'),
feat.MODES))
try:
os.makedirs('out')
except OSError:
pass
print('m[%s]M[%s]' % (np.min(out), np.max(out)))
cv2.imwrite('out/lena_gradient.jpg', np.array(out * 255,
dtype=np.uint8))
def __call__(self, image):
image = imfeat.convert_image(image, [('opencv', 'gray', 8)])
keypoints, descriptors = cv.ExtractSURF(image, None, cv.CreateMemStorage(), (0, 500, 3, 4))
out = []
for ((x, y), laplacian, size, direction, hessian) in keypoints:
out.append({'x': x,
'y': y,
'scale': size,
'orientation': direction,
'sign': laplacian,
'cornerness': hessian})
return out
def test_save_lena(self):
feat = imfeat.LBP()
for x in range(10):
out = feat.make_feature_mask(imfeat.convert_image(
cv2.imread('test_images/lena.jpg'), feat.MODES),
pool_radius=x)
print(out.shape)
out = out.reshape(out.shape[:2])
try:
os.makedirs('out')
except OSError:
pass
cv2.imwrite('out/lena_lbp-%d.jpg' % x, out)
def _predict(self, image):
# Makes the max size length 320
max_side = 320
image = imfeat.convert_image(image, self.MODES)
sz = np.array([image.shape[1], image.shape[0]])
sz = np.array(max_side * sz / float(np.max(sz)), dtype=np.int)
image = cv2.resize(image, tuple(sz))
image = self._make_masks(image)
# Predict using both trees
max_classes1, max_probs1, leaves1, all_probs1 = self.tp.predict(image, leaves=True, all_probs=True)
pred_integrals = convert_leaves_all_probs_pred(image, leaves1, all_probs1, self.tp.num_leaves)
max_classes2, max_probs2, all_probs2 = self.tp2.predict(pred_integrals, all_probs=True)
return max_classes1, max_probs1, leaves1, max_classes2, max_probs2, all_probs2
def _predict(self, image):
# Makes the max size length 320
max_side = 320
image = imfeat.convert_image(image, self.MODES)
sz = np.array([image.shape[1], image.shape[0]])
sz = np.array(max_side * sz / float(np.max(sz)), dtype=np.int)
image = cv2.resize(image, tuple(sz))
image = self._make_masks(image)
# Predict using both trees
max_classes1, max_probs1, leaves1, all_probs1 = self.tp.predict(image, leaves=True, all_probs=True)
pred_integrals = convert_leaves_all_probs_pred(image, leaves1, all_probs1, self.tp.num_leaves)
max_classes2, max_probs2, all_probs2 = self.tp2.predict(pred_integrals, all_probs=True)
return max_classes1, max_probs1, leaves1, max_classes2, max_probs2, all_probs2
def _feature_hist(self, image):
out = []
for block in imfeat.BlockGenerator(image,
imfeat.CoordGeneratorRect,
output_size=(self.sbin, self.sbin),
step_delta=(self.sbin, self.sbin)):
out.append(
imfeat.convert_image(image, {
'type': 'numpy',
'dtype': 'float32',
'mode': 'lab'
}).ravel())
return np.asfarray(out)
def __call__(self, image, num_components=None):
if num_components is None:
num_components = self.num_componenets
image = imfeat.convert_image(image, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'gray'})
input_path = self.temp_dir + '/input.pgm'
output_path = self.temp_dir + '/output.ppm'
#image = Image.fromarray(image)
#image.save(input_path)
cv2.imwrite(input_path, image)
cur_dir = os.path.abspath('.')
os.chdir(self.temp_dir)
cmd = './superpixels input.pgm output.ppm %d' % (num_components,)
subprocess.call(cmd.split())
out = cv2.imread(output_path)
os.chdir(cur_dir)
return out
def test_cvhsv(self):
"""Check that HSV conversion produces correct values for a solid color
test case
"""
# Check that HSV produces a 'blue' hue for a blue image
hsv_target = 120.0
# Build the image in BGR order
blue_image = np.tile([255, 0, 0], (10, 10, 1)).astype("u1")
image = cv.CreateImageHeader((10, 10), cv.IPL_DEPTH_8U, 3)
cv.SetData(image, blue_image.tostring())
mode = ("opencv", "hsv", cv.IPL_DEPTH_8U)
hsv = imfeat.convert_image(image, [mode])
hsv = cvBGRtoarray(hsv)
np.testing.assert_allclose(hsv[:, :, 2], hsv_target)
def test_save_lena(self):
feat = imfeat.HOGLatent(8)
image_input = imfeat.convert_image(
cv2.imread('test_images/mosaic_001_01.jpg'), feat.MODES) #lena.jpg
num_eq = 0
num_neq = 0
try:
os.makedirs('out/latent/')
except OSError:
pass
for x in range(260, 261):
#image_input = imfeat.resize_image(image_input, x, x)
for sz in range(1, 6):
print(sz)
sz = 2**sz
num_blocks = (np.floor(
np.asfarray(image_input.shape[:2]) / float(sz) + .5) - 2)
print(num_blocks)
if any(num_blocks <= 0):
continue
#effective_size = (np.floor(np.asfarray(image_input.shape[:2]) / float(sz) + .5)) * sz
feat = imfeat.HOGLatent(sz)
im = image_input.copy()
out = feat.make_feature_mask(im)
print('Dims[%d]' % out.shape[2])
for i in range(out.shape[2]):
out_s = out[:, :, i]
print(np.min(out_s))
print(np.max(out_s))
print('sz[%s]M[%s] m[%s]' %
(sz, np.max(out_s), np.min(out_s)))
out_s = np.array(
255 * (out_s - np.min(out_s)) /
(np.max(out_s) - np.min(out_s) + .000000001),
dtype=np.uint8)
cv2.imwrite('out/latent/lena-hog-%.3d-%.3d.png' % (sz, i),
out_s)
y, x = np.random.randint(0, im.shape[0]), np.random.randint(
0, im.shape[1])
im[y, x, :] += 100
out2 = feat.make_feature_mask(im)
if out[y, x, :].tolist() == out2[y, x, :].tolist():
num_eq += 1
print('-----------%s' % str((num_eq, num_neq)))
else:
num_neq += 1
def test_patterns(self):
feat = imfeat.LBP()
for in_str, out_val in [('[1,1,1;1,2,1;1,1,1]', 255),
('[1,1,1;1,1,1;1,1,1]', 0),
('[0,1,1;1,1,1;1,1,1]', 1),
('[1,0,1;1,1,1;1,1,1]', 2),
('[1,1,0;1,1,1;1,1,1]', 4),
('[1,1,1;1,1,0;1,1,1]', 8),
('[1,1,1;1,1,1;1,1,0]', 16),
('[1,1,1;1,1,1;1,0,1]', 32),
('[1,1,1;1,1,1;0,1,1]', 64),
('[1,1,1;0,1,1;1,1,1]', 128),
('[0,1,1;0,1,1;1,1,1]', 129)]:
data = np.array(np.mat(in_str).A, dtype=np.uint8)
expected = np.ones((3, 3), dtype=np.uint8) * out_val
out = feat.make_feature_mask(imfeat.convert_image(data, feat.MODES))
np.testing.assert_equal(out, expected)
def test_patterns(self):
feat = imfeat.LBP()
for in_str, out_val in [('[1,1,1;1,2,1;1,1,1]', 255),
('[1,1,1;1,1,1;1,1,1]', 0),
('[0,1,1;1,1,1;1,1,1]', 1),
('[1,0,1;1,1,1;1,1,1]', 2),
('[1,1,0;1,1,1;1,1,1]', 4),
('[1,1,1;1,1,0;1,1,1]', 8),
('[1,1,1;1,1,1;1,1,0]', 16),
('[1,1,1;1,1,1;1,0,1]', 32),
('[1,1,1;1,1,1;0,1,1]', 64),
('[1,1,1;0,1,1;1,1,1]', 128),
('[0,1,1;0,1,1;1,1,1]', 129)]:
data = np.array(np.mat(in_str).A, dtype=np.uint8)
expected = np.ones((3, 3), dtype=np.uint8) * out_val
out = feat.make_feature_mask(imfeat.convert_image(
data, feat.MODES))
np.testing.assert_equal(out, expected)
def __call__(self, image, num_components=None):
if num_components is None:
num_components = self.num_componenets
image = imfeat.convert_image(image, {
'type': 'numpy',
'dtype': 'uint8',
'mode': 'gray'
})
input_path = self.temp_dir + '/input.pgm'
output_path = self.temp_dir + '/output.ppm'
#image = Image.fromarray(image)
#image.save(input_path)
cv2.imwrite(input_path, image)
cur_dir = os.path.abspath('.')
os.chdir(self.temp_dir)
cmd = './superpixels input.pgm output.ppm %d' % (num_components, )
subprocess.call(cmd.split())
out = cv2.imread(output_path)
os.chdir(cur_dir)
return out
def gen():
try:
frame_num = -1
skip_next = None
while True:
frame = _read_ppm(proc.stdout)
frame_num += 1
if frame is None:
break
if skip_next is not None:
if frame_num < skip_next: continue
else:
if frame_num % mod != 0: continue
skip_next = yield(frame_num,
frame_num / fps,
imfeat.convert_image(frame, image_modes))
finally:
# Kill the ffmpeg process early if the generator is destroyed
proc.kill()
proc.wait()
def frame_iter(stream, image_modes, mod=1):
SEEK_START_ATTEMPTS = 3
# Use seek to find the first good frame
for i in range(SEEK_START_ATTEMPTS):
try:
stream.tv.seek_to_frame(i)
except IOError:
continue
else:
break
fps = stream.tv.get_fps()
cnt = 0
while 1:
if cnt % mod == 0:
_, num, frame = stream.tv.get_current_frame()[:3]
yield num, num / fps, imfeat.convert_image(frame, image_modes)
try:
stream.tv.get_next_frame()
except IOError:
break
cnt += 1
def test_save_lena(self):
feat = imfeat.HOGLatent(8)
image_input = imfeat.convert_image(cv2.imread('test_images/mosaic_001_01.jpg'), feat.MODES)#lena.jpg
num_eq = 0
num_neq = 0
try:
os.makedirs('out/latent/')
except OSError:
pass
for x in range(260, 261):
#image_input = imfeat.resize_image(image_input, x, x)
for sz in range(1, 6):
print(sz)
sz = 2 ** sz
num_blocks = (np.floor(np.asfarray(image_input.shape[:2]) / float(sz) + .5) - 2)
print(num_blocks)
if any(num_blocks <= 0):
continue
#effective_size = (np.floor(np.asfarray(image_input.shape[:2]) / float(sz) + .5)) * sz
feat = imfeat.HOGLatent(sz)
im = image_input.copy()
out = feat.make_feature_mask(im)
print('Dims[%d]' % out.shape[2])
for i in range(out.shape[2]):
out_s = out[:, :, i]
print(np.min(out_s))
print(np.max(out_s))
print('sz[%s]M[%s] m[%s]' % (sz, np.max(out_s), np.min(out_s)))
out_s = np.array(255 * (out_s - np.min(out_s)) / (np.max(out_s) - np.min(out_s) + .000000001), dtype=np.uint8)
cv2.imwrite('out/latent/lena-hog-%.3d-%.3d.png' % (sz, i), out_s)
y, x = np.random.randint(0, im.shape[0]), np.random.randint(0, im.shape[1])
im[y, x, :] += 100
out2 = feat.make_feature_mask(im)
if out[y, x, :].tolist() == out2[y, x, :].tolist():
num_eq += 1
print('-----------%s' % str((num_eq, num_neq)))
else:
num_neq += 1
def pil_to_cv(fp):
return imfeat.convert_image(Image.open(fp), [('opencv', 'bgr', 8)])
def _load_cv_image(self, value):
image = Image.open(StringIO.StringIO(value))
image = image.convert('RGB')
return image, imfeat.convert_image(image, [('opencv', 'rgb', 8)])
def _make_masks(self, image):
image = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'bgr', 'dtype': 'uint8'}])
image_gray = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'gray', 'dtype': 'uint8'}])
image_gradient = np.array(self.grad.make_feature_mask(np.array(image_gray, dtype=np.float32) / 255.) * 255, dtype=np.uint8)
image_lab = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'lab', 'dtype': 'uint8'}])
return np.ascontiguousarray(np.dstack([image, image_lab, image_gradient]), dtype=np.uint8)
def feat_func(self, frame):
gray_frame = imfeat.convert_image(frame, [('opencv', 'gray', 8)])
return {
'surf': self._surf(gray_frame),
'hist': self._hist_feat_func(gray_frame)
}
def _feature(self, image):
out = []
for block, coords in imfeat.BlockGenerator(image, imfeat.CoordGeneratorRect, output_size=(self.sbin, self.sbin), step_delta=(self.sbin, self.sbin)):
out.append(imfeat.convert_image(block, {'type': 'numpy', 'dtype': 'float32', 'mode': self.mode}).ravel())
return np.asfarray(out)
def _load_cv_image(self, value):
return imfeat.convert_image(Image.open(StringIO.StringIO(value)),
[('opencv', 'rgb', 8)])
def _make_masks(self, image):
image = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'bgr', 'dtype': 'uint8'}])
image_gray = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'gray', 'dtype': 'uint8'}])
image_gradient = np.array(self.grad.make_feature_mask(np.array(image_gray, dtype=np.float32) / 255.) * 255, dtype=np.uint8)
image_lab = imfeat.convert_image(image, [{'type': 'numpy', 'mode': 'lab', 'dtype': 'uint8'}])
return np.ascontiguousarray(np.dstack([image, image_lab, image_gradient]), dtype=np.uint8)
def test_tb(path):
# test that we can read each file using _load_cv_image
for (key, val) in hadoopy.readtb(path):
print(key)
i = imfeat.convert_image(Image.open(StringIO.StringIO(val)),
[('opencv', 'gray', 8)])
def cv_to_jpg(img):
fp = StringIO.StringIO()
imfeat.convert_image(img, ['RGB']).save(fp, 'JPEG')
fp.seek(0)
return fp.read()
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename), features) where features is a dict
frame_features: List of frame features
file_size: Size in bytes
where each frame feature is a dictionary of
frame_time: Time in seconds
frame_num: Frame number
prev_frame_num: Previous frame number (useful if there is a frame skip)
keyframe: Boolean True/False
surf: List of surf points (see impoint)
face_widths:
face_heights:
predictions: Dictionary of predictions
"""
sys.stderr.write('In Raw:%s\n' % str(event_filename))
print(event_filename)
ext = '.' + event_filename[1].rsplit('.')[1]
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
with self.timer('Writing video data'):
fp.write(video_data)
fp.flush()
kf = keyframe.DecisionTree(min_interval=0)
kf.load()
prev_frame = None
prev_frame_num = 0
all_out = []
sz = len(video_data)
self.timer.start('KF')
try:
for (frame_num, frame_time, frame), iskeyframe in kf(viderator.frame_iter(fp.name,
frozen=True)):
hadoopy.counter('RawFeatures', 'NumFrames')
self.timer.stop('KF')
print(frame_time)
if frame_num > self._max_frames:
break
if frame_num % 100 == 0:
with self.timer('Computing face features'):
faces = _detect_faces(imfeat.convert_image(frame, [('opencv', 'gray', 8)]),
self.cascade)
else:
faces = {}
out = {'frame_time': frame_time, 'frame_num': frame_num,
'prev_frame_num': prev_frame_num, 'keyframe': iskeyframe,
'surf': kf.prev_vec['surf']}
if faces: # If any faces
face_heights = np.array([x[0][3] for x in faces]) / float(frame.height)
face_widths = np.array([x[0][2] for x in faces]) / float(frame.width)
out['face_widths'] = face_widths
out['face_heights'] = face_heights
# Output the cur and previous frames if this is a keyframe
if iskeyframe and np.random.random() < self._frame_output_prob:
out['prev_frame'] = cv_to_jpg(prev_frame)
out['frame'] = cv_to_jpg(frame)
# Compute scene features
with self.timer('Computing scene classifier features'):
frame_res = cv.fromarray(cv2.resize(np.asarray(cv.GetMat(frame)), (self._image_width, self._image_height)))
feature = self._feat(frame_res)
out['predictions'] = dict((classifier_name, classifier.predict(feature))
for classifier_name, classifier in self._classifiers)
# Output JPEG with match lines from the SURF feature
if np.random.random() < self._match_line_prob and prev_frame:
out['surf_image'] = cv_to_jpg(plot_matches(prev_frame, kf.surf_debug['matches'], kf.surf_debug['points0'],
kf.surf_debug['points1'], max_feat_width=kf.max_feat_width))
# Output data buffer
all_out.append(out)
if len(all_out) >= self._block_size:
with self.timer('Yield'):
yield event_filename, {'frame_features': all_out,
'file_size': sz}
all_out = []
prev_frame = frame
prev_frame_num = frame_num
self.timer.start('KF')
except viderator.FPSParseException: # NOTE(brandyn): This will disregard videos with this error
hadoopy.counter('SkippedVideos', 'FPSParseException')
return
if all_out:
with self.timer('Yield'):
yield event_filename, {'frame_features': all_out,
'file_size': sz}
def __call__(self, image):
image_cv = imfeat.convert_image(image, self.MODES)
return self.make_features(image_cv)[0]
def imread(self, fn):
out = cv2.imread(fn)
if out is not None:
return out
return imfeat.convert_image(Image.open(fn), {'type': 'numpy', 'dtype': 'uint8', 'mode': 'bgr'})
def __init__(self, images=None, vectors=None, verbose=False):
self.MODES = [('opencv', 'gray', 32)]
self.verbose = verbose
if not images == None:
images = [imfeat.convert_image(i, self.MODES) for i in images]
self.train(images, vectors)
def __call__(self, image):
image_cv = imfeat.convert_image(image, self.MODES)
return self.make_features(image_cv)[0]
def __init__(self, images=None, vectors=None, verbose=False):
self.MODES = [('opencv', 'gray', 32)]
self.verbose = verbose
if not images == None:
images = [imfeat.convert_image(i, self.MODES) for i in images]
self.train(images, vectors)
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename), features) where features is a dict
frame_features: List of frame features
file_size: Size in bytes
where each frame feature is a dictionary of
frame_time: Time in seconds
frame_num: Frame number
prev_frame_num: Previous frame number (useful if there is a frame skip)
keyframe: Boolean True/False
surf: List of surf points (see impoint)
face_widths:
face_heights:
predictions: Dictionary of predictions
"""
sys.stderr.write('In Raw:%s\n' % str(event_filename))
print(event_filename)
ext = '.' + event_filename[1].rsplit('.')[1]
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
with self.timer('Writing video data'):
fp.write(video_data)
fp.flush()
kf = keyframe.DecisionTree(min_interval=0)
kf.load()
prev_frame = None
prev_frame_num = 0
all_out = []
sz = len(video_data)
self.timer.start('KF')
try:
for (frame_num, frame_time, frame), iskeyframe in kf(
viderator.frame_iter(fp.name, frozen=True)):
hadoopy.counter('RawFeatures', 'NumFrames')
self.timer.stop('KF')
print(frame_time)
if frame_num > self._max_frames:
break
if frame_num % 100 == 0:
with self.timer('Computing face features'):
faces = _detect_faces(
imfeat.convert_image(frame,
[('opencv', 'gray', 8)]),
self.cascade)
else:
faces = {}
out = {
'frame_time': frame_time,
'frame_num': frame_num,
'prev_frame_num': prev_frame_num,
'keyframe': iskeyframe,
'surf': kf.prev_vec['surf']
}
if faces: # If any faces
face_heights = np.array([x[0][3] for x in faces
]) / float(frame.height)
face_widths = np.array([x[0][2] for x in faces
]) / float(frame.width)
out['face_widths'] = face_widths
out['face_heights'] = face_heights
# Output the cur and previous frames if this is a keyframe
if iskeyframe and np.random.random(
) < self._frame_output_prob:
out['prev_frame'] = cv_to_jpg(prev_frame)
out['frame'] = cv_to_jpg(frame)
# Compute scene features
with self.timer('Computing scene classifier features'):
frame_res = cv.fromarray(
cv2.resize(
np.asarray(cv.GetMat(frame)),
(self._image_width, self._image_height)))
feature = self._feat(frame_res)
out['predictions'] = dict(
(classifier_name, classifier.predict(feature)) for
classifier_name, classifier in self._classifiers)
# Output JPEG with match lines from the SURF feature
if np.random.random(
) < self._match_line_prob and prev_frame:
out['surf_image'] = cv_to_jpg(
plot_matches(prev_frame,
kf.surf_debug['matches'],
kf.surf_debug['points0'],
kf.surf_debug['points1'],
max_feat_width=kf.max_feat_width))
# Output data buffer
all_out.append(out)
if len(all_out) >= self._block_size:
with self.timer('Yield'):
yield event_filename, {
'frame_features': all_out,
'file_size': sz
}
all_out = []
prev_frame = frame
prev_frame_num = frame_num
self.timer.start('KF')
except viderator.FPSParseException: # NOTE(brandyn): This will disregard videos with this error
hadoopy.counter('SkippedVideos', 'FPSParseException')
return
if all_out:
with self.timer('Yield'):
yield event_filename, {
'frame_features': all_out,
'file_size': sz
}
def _load_cv_image(self, value):
image = Image.open(StringIO.StringIO(value))
image = image.convert('RGB')
return image, imfeat.convert_image(image, [('opencv', 'rgb', 8)])
def _load_cv_image(self, value):
return imfeat.convert_image(Image.open(StringIO.StringIO(value)),
[('opencv', 'bgr', 8)])
def cv_to_jpg(img):
fp = StringIO.StringIO()
imfeat.convert_image(img, ['RGB']).save(fp, 'JPEG')
fp.seek(0)
return fp.read()
def convert_color(image):
image = imfeat.convert_image(image, [('opencv', 'lab', cv.IPL_DEPTH_8U)])
image = np.asarray(cv.GetMat(image)).copy()
return image
def feat_func(self, frame):
gray_frame = imfeat.convert_image(frame, [('opencv', 'gray', 8)])
return {'surf': self._surf(gray_frame), 'hist': self._hist_feat_func(gray_frame)}
