def erode_data(data, footprint = None, structure = None):
if footprint is not None:
eroded_data = grey_erosion(data, footprint = footprint, structure = structure)
else:
eroded_data = grey_erosion(data, size=(10, 10))
return eroded_data
def erode_data(data, footprint = None, structure = None):
if footprint is not None:
eroded_data = grey_erosion(data, footprint = footprint, structure = structure)
else:
eroded_data = grey_erosion(data, size=(10, 10))
return eroded_data
def extract_fundamental(amplitude):
fundamental = np.zeros(amplitude.shape)
# TODO: replace all of this with real code or at least clean it up
# it should just be one big numpy thingy
f_band_min = -4
f_band_max = 8
f_band_len = f_band_max - f_band_min
f_band_coeffs = (1 - np.concatenate(
(np.array(range(f_band_min, 0)) / f_band_min,
np.array(range(f_band_max)) / f_band_max)))[:, np.newaxis]
peak_finder = np.array([-0.5, -0.5, 2, -0.5, -0.5])[:, np.newaxis].T
console.time("big loop")
freqs = np.argmax(np.mean(amplitude[:50], axis=2), axis=0)
# console.stats(freqs)
for t in range(amplitude.shape[1]):
f = freqs[t]
# handle case where 2nd harmonic > first
if np.mean(amplitude[f // 2, t]) > 0.4 * np.mean(amplitude[f, t]):
f = f // 2
freqs[t] = f
if f > 5:
f_min = f + f_band_min
f_max = f + f_band_max
fundamental[f_min:f_max,
t] = f_band_coeffs * amplitude[f_min:f_max, t]
console.timeEnd("big loop")
console.time("remove dots")
mask = (grey_dilation(grey_erosion(fundamental,
structure=np.ones((3, 5, 1))),
structure=np.ones((6, 12, 1))) > 0.1)
console.timeEnd("remove dots")
fundamental *= mask
return fundamental
def get_sibilants(content_amplitude, content_fundamental_amps):
num_freqs, num_timesteps, _ = content_amplitude.shape
output = content_amplitude.copy()
clipped_amps = content_fundamental_amps.copy()
clipped_amps[clipped_amps < 0.5] = 0
output -= 2 * clipped_amps[np.newaxis, :, np.newaxis]
output = np.clip(output, 0, 1) # sigh
output = scipy.ndimage.filters.gaussian_filter1d(output,
4,
axis=0,
mode="nearest")
clipped_output = np.clip(output, 0, 1)
thresh = 0.3
clipped_output[output > thresh] = 1
clipped_output[output <= thresh] = 0
clipped_output = grey_erosion(
(clipped_output * 255), structure=np.ones((32, 1, 1))) / 255.0
# TODO: these parts are hacky :(
clipped_output[:300] = 0
output *= clipped_output
output[output > 0.1] *= 4
output = np.clip(output, 0, 1)
output = scipy.ndimage.filters.gaussian_filter1d(output,
128,
axis=0,
mode="nearest")
output = np.sqrt(output)
return output
def _grey_erosion(self):
vol_name = str(self.out_edit.text())
num = self.size_combo.currentIndex() + 3
size = (num, num, num)
mode = self.mode_combo.currentText()
cval = self.cval_edit.text()
if not vol_name:
self.out_edit.setFocus()
return
try:
cval = int(cval)
except ValueError:
self.cval_edit.selectAll()
return
if cval > 255 or cval < 0:
print "cval must be 0-255!"
return
source_row = self.source_combo.currentIndex()
source_data = self._model.data(self._model.index(source_row),
Qt.UserRole + 5)
new_vol = morphology.grey_erosion(source_data,
size=size,
mode=mode,
cval=cval)
self._model.addItem(new_vol, None, vol_name,
self._model._data[0].get_header())
self.done(0)
def post_process(self):
"""
Generate the metrics from the stored raw ray trace results.
"""
data = self.traverse_results
## Check raw traverse results for continguous regions of no ground coverage.
none_runs = [
len(list(grp)) for k, grp in itertools.groupby(data, lambda x: x)
if k is None
]
if none_runs and max(none_runs) >= self.contiguous:
## Fail because there is a contiguous region larger than allowed
fwd = "Fail"
aft = "Fail"
no_fire_area = "Fail"
else:
trav_cen = get_translation(self.weapon.trans_trav)
angles = np.array([x[0] for x in data if x[1] is not None])
muz_pts = np.array([x[1][0] for x in data if x[1] is not None])
gnd_pts = np.array([x[1][1] for x in data if x[1] is not None])
## Distance along the ground
gnd_dist = np.sqrt((trav_cen[0] - gnd_pts[:, 0])**2 +
(trav_cen[2] - gnd_pts[:, 2])**2)
## Calculate the minimum shoot distance over contiguous regions
mod_dist = grey_erosion(gnd_dist, self.contiguous, mode="wrap")
## Calculate an area metric (A = 0.5 * a.b.sin(c_ang))
num_tris = len(angles)
no_fire_area = 0.0
for i in xrange(num_tris):
next_ang = (angles[0] +
2 * pi) if i + 1 >= num_tris else angles[i + 1]
c_ang = next_ang - angles[i]
a = gnd_dist[i]
b = gnd_dist[(i + 1) % num_tris]
no_fire_area += 0.5 * a * b * sin(c_ang)
## Calculate min forward and aft shoot distance.
fwd = np.max(mod_dist[np.where((angles >= 3 * pi / 2)
| (angles <= pi / 2))])
aft = np.max(mod_dist[np.where((angles <= 3 * pi / 2)
& (angles >= pi / 2))])
self.post_process_2d(angles, gnd_dist, mod_dist, no_fire_area, fwd,
aft)
if self.show_3d:
self.post_process_3d(muz_pts, gnd_pts)
## Write results to json file and echo to log.
out = {
"min_fire_dist_fore_180": fwd,
"min_fire_dist_aft_180": aft,
"no_fire_area": no_fire_area
}
tba.write_results(out)
def jeff_coral_finder(im, sand_intensity_threshold, coral_gradient_threshold,
maximum_altseqfilt_radius, shadow_discriminant_threshold,
shadow_discriminant_scaling):
im_grey = N.asarray(im.convert("L"))
im = N.asarray(im)
dot = N.array([[0,1,0], [1,1,1], [0,1,0]])
dilated = morphology.grey_dilation(im_grey, dot.shape, structure=dot)
eroded = morphology.grey_erosion(im_grey, dot.shape, structure=dot)
gradient = dilated - eroded
fisher_discriminant = N.dot(im, shadow_discriminant_scaling)
# Make initial class determinations.
is_shadow = fisher_discriminant < shadow_discriminant_threshold
is_sand = im_grey > sand_intensity_threshold
is_smooth = gradient < coral_gradient_threshold
is_coral = is_smooth & ~is_sand & ~is_shadow
# Now perform an alternating sequence filter on coral,
for radius in range(1, maximum_altseqfilt_radius+1):
se = disk_strel(radius)
opened = morphology.binary_opening(is_coral, se)
is_coral = morphology.binary_closing(opened, se)
# Now perform an alternating sequence filter on sand.
for radius in range(1, maximum_altseqfilt_radius+1):
se = disk_strel(radius)
opened = morphology.binary_opening(is_sand, se)
is_sand = morphology.binary_closing(opened, se)
# Use coral mask to exclude sand.
is_sand = is_sand & ~is_coral
return is_sand, is_coral
def prepare_data_fgbg_weigthed(batch_x,
batch_y,
border_weight=2,
separation_border_weight=5,
sigma=1):
# Wrap x in np array and add channel dimension
out_x = np.array(batch_x)[..., None] # TODO input with channels?
# Create the weight map
struct = morphology.generate_binary_structure(len(batch_y[0].shape), 1)
foreground = np.zeros_like(batch_y)
weight_map = np.zeros_like(batch_y, dtype='float32')
for i, mask in enumerate(batch_y):
borders = morphology.morphological_laplace(mask, structure=struct) \
> (np.max(mask) + 1)
separation_borders = np.logical_and(
morphology.grey_erosion(mask, structure=struct), borders)
weight_map[i] = separation_border_weight * separation_borders \
+ border_weight * borders \
+ 1
# Filter weight map
if sigma > 0:
weight_map[i] = filters.gaussian_filter(weight_map[i], sigma=sigma)
# Foreground is the mask without the borders
foreground[i] = np.logical_and((mask > 0), np.logical_not(borders))
background = np.logical_not(foreground)
out_y = np.array(np.stack([foreground, background], axis=-1),
dtype='float32')
return [out_x, weight_map], out_y
def erosion(parameters):
"""Erodes a greyscale image.
For the simple case of a full and flat structuring element, it can be
viewed as a minimum filter over a sliding window.
It wraps `scipy.ndimage.morphology.grey_erosion`. The `footprint`,
`structure`, `output`, `mode`, `cval` and `origin` options are not
supported.
Keep in mind that `mode` and `cval` influence the results. In this case
the default mode is used, `reflect`.
:param parameters['data'][0]: input array
:type parameters['data'][0]: numpy.array
:param parameters['size']: which neighbours to take into account, defaults
to (3, 3) a.k.a. numpy.ones((3, 3))
:type parameters['size']: list
:return: numpy.array
"""
data = parameters['data'][0]
size = tuple(parameters['size'])
return morphology.grey_erosion(data, size=size)
def _grey_erosion(self):
vol_name = str(self.out_edit.text())
num = self.size_combo.currentIndex() + 3
size = (num, num, num)
mode = self.mode_combo.currentText()
cval = self.cval_edit.text()
if not vol_name:
self.out_edit.setFocus()
return
try:
cval = int(cval)
except ValueError:
self.cval_edit.selectAll()
return
if cval>255 or cval<0:
print "cval must be 0-255!"
return
source_row = self.source_combo.currentIndex()
source_data = self._model.data(self._model.index(source_row),
Qt.UserRole + 5)
new_vol = morphology.grey_erosion(source_data,size=size,mode=mode,cval=cval)
self._model.addItem(new_vol,
None,
vol_name,
self._model._data[0].get_header())
self.done(0)
def _adapt(self, video_idx, frame_idx, last_mask, get_posteriors_fn):
eroded_mask = grey_erosion(last_mask, size=(self.erosion_size, self.erosion_size, 1))
dt = distance_transform_edt(numpy.logical_not(eroded_mask))
adaptation_target = numpy.zeros_like(last_mask)
adaptation_target[:] = VOID_LABEL
current_posteriors = get_posteriors_fn()
positives = current_posteriors[:, :, 1] > self.posterior_positive_threshold
if self.use_positives:
adaptation_target[positives] = 1
threshold = self.distance_negative_threshold
negatives = dt > threshold
if self.use_negatives:
adaptation_target[negatives] = 0
do_adaptation = eroded_mask.sum() > 0
if self.debug:
adaptation_target_visualization = adaptation_target.copy()
adaptation_target_visualization[adaptation_target == 1] = 128
if not do_adaptation:
adaptation_target_visualization[:] = VOID_LABEL
from scipy.misc import imsave
folder = self.val_data.video_tag().replace("__", "/")
imsave("forwarded/" + self.model + "/valid/" + folder + "/adaptation_%05d.png" % frame_idx,
numpy.squeeze(adaptation_target_visualization))
self.train_data.set_video_idx(video_idx)
for idx in range(self.n_adaptation_steps):
do_step = True
if idx % self.adaptation_interval == 0:
if do_adaptation:
feed_dict = self.train_data.feed_dict_for_video_frame(frame_idx, with_annotations=True)
feed_dict[self.train_data.get_label_placeholder()] = adaptation_target
loss_scale = self.adaptation_loss_scale
adaption_frame_idx = frame_idx
else:
print("skipping current frame adaptation, since the target seems to be lost", file=log.v4)
do_step = False
else:
# mix in first frame to avoid drift
# (do this even if we think the target is lost, since then this can help to find back the target)
feed_dict = self.train_data.feed_dict_for_video_frame(frame_idx=0, with_annotations=True)
loss_scale = 1.0
adaption_frame_idx = 0
if do_step:
loss, _, n_imgs = self.trainer.train_step(epoch=idx, feed_dict=feed_dict, loss_scale=loss_scale,
learning_rate=self.adaptation_learning_rate)
assert n_imgs == 1
print("adapting on frame", adaption_frame_idx, "of sequence", video_idx + 1, \
self.train_data.video_tag(video_idx), "loss:", loss, file=log.v4)
if do_adaptation:
return negatives
else:
return None
def filtering(img, filter_type="gaussian", filter_value=0.5):
"""
:Parameters:
- `image` (openalea.image.SpatialImage) - image
- `filter_type`(str) - denoising method used for filtering ("gaussian" or "asf" for alternate sequential filter).
default is "gaussian".
- `filter_value` (float for "gaussian" filter or int for alternate sequential filter) - value used for the filtering :
* for a Gaussian filtering, the "filter_value" corresponds to the standard deviation.
* for a Alternate Sequential Filter, the "filter_value" corresponds to the number of
succession of morphological opening and closing operations.
"""
# if not isinstance(img, SpatialImage):
# img = SpatialImage(img)
#
# if filter_type == 'gaussian':
# if not isinstance(filter_value, float):
# raise RuntimeError, 'value used for Gaussian filtering must be a float type'
# else:
# img = recfilters(img, filter_type="sigma", filter_value=filter_value, Trueyz=(0, 0, 0))
if filter_type == 'asf':
if not isinstance(filter_value, int):
raise RuntimeError(
'value used for the Alternate Sequential Filter must be a integer type'
)
else:
for rad in range(1, ((filter_value + 1) / 2) + 1):
print(
"closing operations with structuring elements of size %s" %
rad)
struct = euclidean_sphere(rad)
# ~ s=(rad,rad,rad)
img = grey_dilation(img, footprint=struct)
img = grey_erosion(img, footprint=struct)
if filter_value >= rad * 2:
print(
"opening operations with structuring elements of size %s"
% rad)
img = grey_erosion(img, footprint=struct)
img = grey_dilation(img, footprint=struct)
else:
raise RuntimeError('filter type not supported')
return img
def ruined_pic(pic):
''' alter the number of pixels in the image: increase or decrease character width '''
val = random.randint(1, 2)
if val == 1:
pic = grey_dilation(pic.reshape(64, 64), size=(2, 1))
else:
pic = grey_erosion(pic.reshape(64, 64), size=(2, 1))
return pic
def threshold(g,sigma,frac,nbins,erosion_diameter):
g = background_subtract(g,strel())
g = gaussian_blur(g,sigma)
g = np.round(g)
gthreshold = get_poisson_threshold(g,frac=frac,nbins=nbins)
gt = g.copy()
gt[np.where(g<gthreshold)] = 0.0
gto = morphology.grey_erosion(gt,footprint=strel(diameter=erosion_diameter))
return gto
def post_process(self):
"""
Generate the metrics from the stored raw ray trace results.
"""
data = self.traverse_results
## Check raw traverse results for continguous regions of no ground coverage.
none_runs = [len(list(grp)) for k, grp in itertools.groupby(data, lambda x:x) if k is None]
if none_runs and max(none_runs) >= self.contiguous:
## Fail because there is a contiguous region larger than allowed
fwd = "Fail"
aft = "Fail"
no_fire_area = "Fail"
else:
trav_cen = get_translation(self.weapon.trans_trav)
angles = np.array([x[0] for x in data if x[1] is not None])
muz_pts = np.array([x[1][0] for x in data if x[1] is not None])
gnd_pts = np.array([x[1][1] for x in data if x[1] is not None])
## Distance along the ground
gnd_dist = np.sqrt((trav_cen[0] - gnd_pts[:, 0]) ** 2 +
(trav_cen[2] - gnd_pts[:, 2]) ** 2)
## Calculate the minimum shoot distance over contiguous regions
mod_dist = grey_erosion(gnd_dist, self.contiguous, mode="wrap")
## Calculate an area metric (A = 0.5 * a.b.sin(c_ang))
num_tris = len(angles)
no_fire_area = 0.0
for i in xrange(num_tris):
next_ang = (angles[0] + 2 * pi) if i + 1 >= num_tris else angles[i+1]
c_ang = next_ang - angles[i]
a = gnd_dist[i]
b = gnd_dist[(i + 1) % num_tris]
no_fire_area += 0.5 * a * b * sin(c_ang)
## Calculate min forward and aft shoot distance.
fwd = np.max(mod_dist[np.where((angles >= 3 * pi / 2) | (angles <= pi / 2))])
aft = np.max(mod_dist[np.where((angles <= 3 * pi / 2) & (angles >= pi / 2))])
self.post_process_2d(angles, gnd_dist, mod_dist, no_fire_area, fwd, aft)
if self.show_3d:
self.post_process_3d(muz_pts, gnd_pts)
## Write results to json file and echo to log.
out = {
"min_fire_dist_fore_180" : fwd,
"min_fire_dist_aft_180" : aft,
"no_fire_area" : no_fire_area
}
tba.write_results(out)
def extract_watermark(audio_file, interactive=False):
"""
Extracts the watermark from the spectrogram of the given audio file
:param audio_file: path to wav file
:param interactive: activates plotting
:return: watermark as text, or None if the watermark could not be extracted
"""
# Convert audio file to wav if necessary
wavFile = convert_to_wav(audio_file)
fs, data = wavfile.read(wavFile)
data = data.astype(np.float) / np.max(np.abs(data))
window_length = 1024
nfft = window_length
h = window_length // 4
spectrogram, f, t = stft(data, window_length, h, nfft, fs)
if interactive:
plot_spectrogram(spectrogram)
# Convert to PIL image in order to use optical character recognition
# Flip upside down due to the usual way in which we view a spectrogram
ocr_image = np.flipud(np.abs(spectrogram))
# Do some image enhancement
ocr_image[ocr_image < 0.2] = 0
ocr_image = grey_closing(ocr_image, (5, 2))
ocr_image = grey_erosion(ocr_image, (3, 5))
# Convert to 8 bit image
ocr_image = np.uint8(ocr_image / np.max(ocr_image) * 255 * 10)[20:120, :]
ocr_image[ocr_image > 5] = 255
# Enlarge image by interpolation
# ocr_image = imresize(ocr_image, (ocr_image.shape[0] * 8, ocr_image.shape[1] * 8), interp="bilinear")
if interactive:
# Show for debugging purposes
plt.imshow(ocr_image)
plt.show()
ocr_image = Image.fromarray(ocr_image)
ocr_image_filename = "test.png"
ocr_image.save(ocr_image_filename, format="png")
# watermark = ocr.tesseract(ocr_image)
watermark = ocr_space(ocr_image_filename)
# ocr_image.save("test.png", format="png")
return watermark
def dist_to_dep(dist_maps, cam_Ks, **kwargs):
'''
transform distance maps to depth maps.
:param dist_maps: distance value maps from camera to poins
:param cam_Ks: camera intrinsics
:return: depth maps: z values from camera to points.
'''
depth_maps = np.ones_like(dist_maps) * np.inf
view_id = 0
for dist_map, cam_K in zip(dist_maps, cam_Ks):
u, v = np.meshgrid(range(dist_map.shape[1]), range(dist_map.shape[0]))
u = u.reshape([1, -1])[0]
v = v.reshape([1, -1])[0]
dist_cam = dist_map[v, u]
non_inf_indices = np.argwhere(dist_cam<np.inf).T[0]
dist_cam = dist_cam[non_inf_indices]
u = u[non_inf_indices]
v = v[non_inf_indices]
# calculate coordinates
x_temp = (u - cam_K[0][2]) / cam_K[0][0]
y_temp = (v - cam_K[1][2]) / cam_K[1][1]
z_temp = 1
z = dist_cam / np.sqrt(x_temp**2 + y_temp**2 + z_temp**2)
depth_maps[view_id, v, u] = z
if 'erosion_size' in kwargs:
# This is mainly result of experimenting.
# The core idea is that the volume of the object is enlarged slightly
# (by subtracting a constant from the depth map).
# Dilation additionally enlarges thin structures (e.g. for chairs). (refers to Occupancy Network)
depth_maps[view_id] = grey_erosion(depth_maps[view_id],
size=(kwargs['erosion_size'], kwargs['erosion_size']))
view_id += 1
return depth_maps
def get_correct_flat_and_mask(self, flat_filename, ass_filename, pos_filename):
"""usage: get_correct_flat_and_mask(flat_filename, ass_filename, pos_filename)
"""
calib_dir = self.main_dir + '/calib/'
# Reading flat, assign, and position files
image_flat = MioDataRedProc(calib_dir + flat_filename)
print('Reading ' + flat_filename + ' completed!')
assf = mf.ReadASSFile(calib_dir + ass_filename)
self.arm_assign_list = assf.get_armcoor()
posf = mf.ReadPOSFile(calib_dir + pos_filename)
self.fibre_pos_list = posf.get_position()
# creating two new dictionary with the same keys of "arm_assign_list"
self.active_fibre = self.arm_assign_list.keys()
fibre_mask = self.arm_assign_list.fromkeys(self.active_fibre) # fibre mask
correct_flat = self.arm_assign_list.fromkeys(self.active_fibre)
# calling Bias Rigions 6 images. Just need to call once from the same flat or data images.
print('and obtaining flat bias corners')
br6_flat = image_flat.BR6Image()
# correcting flat images and create an averaged flat template
for fibre_number in self.active_fibre:
x, y = self.fibre_pos_list[fibre_number] # x, y positions of fibre image
# subtracting instead bias level to obtain a corrected flat iamge
correct_flatimage = image_flat.FibreImage(x,y) - br6_flat
average_flat = np.mean(correct_flatimage, axis=0) # creating an averaged flat
# normalising flat image
correct_flat[fibre_number] = average_flat / np.amax(average_flat)
# creating a fibre mask to remove the non_data edge pixel
mask_erosion = morphology.grey_erosion(average_flat, size=(1,1))
fibre_mask[fibre_number] = mask_erosion > np.median(mask_erosion) - 1.5*np.std(mask_erosion)
return correct_flat, fibre_mask
def build_iscat_training(bf_filepaths, iscat_filepaths, sampling=4):
"""Creates iscat training data and target in data/iscat_seg/[REF_FRAMES / MASKS] for the iSCAT cell segmentation task
ARGS:
bf_filepaths (list(str)): filepaths of all the bright field images to input as returned by utilitiues.load_data_paths()
iscat_filepaths (list(str)): filepaths of all the iscat images to input as returned by utilitiues.load_data_paths()
sampling (int): sampling interval of the saved images (lower storage footprint)
"""
OUT_PATH = DATA_PATH + 'iscat_seg/'
os.makedirs(os.path.join(OUT_PATH, 'REF_FRAMES/'), exist_ok=True)
os.makedirs(os.path.join(OUT_PATH, 'MASKS/'), exist_ok=True)
# Range of non filtered elements [px]
min_size, max_size = 1, 13
iscat_stacks = (utilities.load_imgs(path) for path in iscat_filepaths)
bf_stacks = (utilities.load_imgs(path) for path in bf_filepaths)
# Returns the metadata of the exwperiments such as frame rate
metadatas = get_experiments_metadata(iscat_filepaths)
if torch.cuda.is_available():
device = torch.cuda.current_device()
torch.cuda.set_device(device)
print("Running on: {:s}".format(torch.cuda.get_device_name(device)))
cuda = torch.device('cuda')
else:
# Doesn't run on CPU only machines comment if no GPU
print("No CUDA device found")
sys.exit(1)
unet = UNetCell(1, 1, device=cuda, bilinear_upsampling=False)
unet.load_state_dict(torch.load('outputs/saved_models/bf_unet.pth'))
for i, (bf_stack, iscat_stack,
metadata) in enumerate(zip(bf_stacks, iscat_stacks, metadatas)):
if i < 45: continue
bf_stack = bf_stack.astype('float32')
print(bf_stack.shape)
if bf_stack.shape[1:] != iscat_stack.shape[1:]:
bf_stack = processing.coregister(bf_stack, 1.38)
print(bf_stack.shape)
normalize(bf_stack)
# Samples iscat image to correct for the difference in framefate
iscat_stack = iscat_stack[::sampling * int(metadata['iscat_fps'] /
metadata['tirf_fps'])]
torch_stack = torch.from_numpy(bf_stack).unsqueeze(1).cuda()
mask = unet.predict_stack(
torch_stack).detach().squeeze().cpu().numpy() > 0.05
mask = morphology.grey_erosion(mask * 255,
structure=processing.structural_element(
'circle', (3, 5, 5)))
mask = morphology.grey_closing(mask,
structure=processing.structural_element(
'circle', (3, 7, 7)))
mask = (mask > 50).astype('uint8')
# Median filtering and normalization
iscat_stack = processing.image_correction(iscat_stack)
# Contrast enhancement
iscat_stack = processing.enhance_contrast(iscat_stack,
'stretching',
percentile=(1, 99))
# Fourier filtering of image
iscat_stack = processing.fft_filtering(iscat_stack, min_size, max_size,
True)
iscat_stack = processing.enhance_contrast(iscat_stack,
'stretching',
percentile=(3, 97))
for j in range(0, min(iscat_stack.shape[0], mask.shape[0]), sampling):
if iscat_stack[j].shape == mask[j].shape:
# Doesn't save images without detected cells
if mask[j].max() == 0: continue
print("\rSaving to stack_{}_{}.png".format(i + 1, j + 1),
end=' ' * 5)
tifffile.imsave(
os.path.join(OUT_PATH, 'REF_FRAMES/',
"stack_{}_{}.png".format(i + 1, j + 1)),
rescale(iscat_stack[j]))
tifffile.imsave(
os.path.join(OUT_PATH, 'MASKS/',
"mask_{}_{}.png".format(i + 1, j + 1)),
mask[j] * 255)
else:
print("Error, shape: {}, {}".format(iscat_stack[j].shape,
mask[j].shape))
break
print('')
def _adapt(self, video_idx, frame_idx, last_mask, get_posteriors_fn,
kframe_lst):
eroded_mask = grey_erosion(last_mask,
size=(self.erosion_size, self.erosion_size,
1))
adaptation_target2 = last_mask
adaptation_target = np.zeros_like(last_mask)
adaptation_target[:] = VOID_LABEL
current_posteriors = get_posteriors_fn()
positives = current_posteriors[:, :,
1] > self.posterior_positive_threshold
if self.use_positives:
adaptation_target[positives] = 1
dt = distance_transform_edt(np.logical_not(eroded_mask))
threshold = self.distance_negative_threshold
negatives = dt > threshold
if self.use_negatives:
adaptation_target[negatives] = 0
do_adaptation = eroded_mask.sum() > 0
print('frame_idx', frame_idx)
#if self.debug:
# adaptation_target_visualization = adaptation_target.copy()
# adaptation_target_visualization[adaptation_target == 1] = 128
# if not do_adaptation:
# adaptation_target_visualization[:] = VOID_LABEL
# from scipy.misc import imsave
# folder = self.val_data.video_tag().replace("__", "/")
# imsave("forwarded/" + self.model + "/valid/" + folder + "/adaptation_%05d.png" % frame_idx,
# np.squeeze(adaptation_target_visualization))
self.train_data.set_video_idx(video_idx)
threshold_ = 0.05
#for idx in xrange(self.n_adaptation_steps): #
for idx in xrange(frame_idx):
do_step = True
#print(kframe_lst)
if idx % len(kframe_lst) == 0: #adaptation_interval
if do_adaptation:
#print("NewIter")
#print("idx % self.adaptation_interval == 0",idx % self.adaptation_interval == 0)
feed_dict = self.train_data.feed_dict_for_video_frame(
frame_idx, with_annotations=True)
feed_dict[self.train_data.get_label_placeholder(
)] = adaptation_target #
loss_scale = self.adaptation_loss_scale * 5
adaption_frame_idx = frame_idx
else:
print >> log.v4, "skipping current frame adaptation, since the target seems to be lost"
do_step = False
elif idx % len(kframe_lst) == 1:
if len(kframe_lst) == 2:
feed_dict = self.train_data.feed_dict_for_video_frame(
kframe_lst[1], with_annotations=True)
loss_scale = self.adaptation_loss_scale * 10
adaption_frame_idx = kframe_lst[1]
else:
feed_dict = self.train_data.feed_dict_for_video_frame(
kframe_lst[-1], with_annotations=True)
loss_scale = self.adaptation_loss_scale * 10
adaption_frame_idx = kframe_lst[-1]
elif idx % len(kframe_lst) == 2:
#print "----------------2------------"
#print "len kframe",len(kframe_lst)
if len(kframe_lst) > 2:
feed_dict = self.train_data.feed_dict_for_video_frame(
kframe_lst[-2], with_annotations=True)
loss_scale = self.adaptation_loss_scale * 10
adaption_frame_idx = kframe_lst[-2]
else:
# mix in first frame to avoid drift
# (do this even if we think the target is lost, since then this can help to find back the target)
feed_dict = self.train_data.feed_dict_for_video_frame(
0, with_annotations=True)
loss_scale = 1.0
adaption_frame_idx = 0
if do_step:
#self._finetune(video_idx, n_finetune_steps=5)
loss, measures, n_imgs = self.trainer.train_step(
epoch=idx,
feed_dict=feed_dict,
loss_scale=loss_scale,
learning_rate=self.adaptation_learning_rate)
#iou=Measures.calc_iou(measures,n_imgs,[0])
assert n_imgs == 1
#print >> log.v4, "adapting on frame", adaption_frame_idx, "of sequence", video_idx + 1, \
# self.train_data.video_tag(video_idx), "loss:", loss,"iou:",iou
if do_adaptation:
return negatives
else:
return None
def shadows(horz_data, slp, asp, sza, eff_sza, saa):
"""Calculate self, cast and total shadows from a DEM.
Args:
horz_data (ndarray): horizon elevation data computed from the DEM
slp (ndarray): slope calculated from the DEM
asp (ndarray): aspect calculated from the DEM
sza (ndarray): solar zenith angles gridded as horz_data
eff_sza (ndarray): effective solar zenith angles gridded as horz_data
saa (ndarray): solar azimuth angles gridded as horz_data
Returns:
b (ndarray): combined cast and self shadow map. binary product.
(1 = shadow, 0 = no shadow).
bs (ndarray): self shadows. Binary product.
bc (ndarray): cast shadows. binary product."""
# get number of horizon directions from the horizon file
N = horz_data.shape[0]
# Switch horizon data elevation to angle
horz_data = 90 - horz_data
# Calculate self shadows (cos gamma is negative)
bs = np.ones(shape=np.shape(eff_sza))
# get around error due to Nans
eff_sza_nonan = np.copy(eff_sza)
eff_sza_nonan[np.isnan(eff_sza_nonan)] = 1
# Self-shadows with relaxed (slightly positive angle) value
bs[np.cos(eff_sza_nonan) < 0.035] = 0
# Elementary angle between horizon lines
dphi = 2 * np.pi / N
# Find the horizon line surrounding a given azimuth
nn1 = np.int8(np.floor(saa / dphi))
nn2 = np.int8(np.ceil(saa / dphi))
m1 = np.uint32(np.mod(N / 2 - nn1, N) + 1)
m2 = np.uint32(np.mod(N / 2 - nn2, N) + 1)
m1prodshape = (np.shape(m1)[0] * np.shape(m1)[1])
m1L = m1prodshape * (m1.flatten() - 1) + np.uint32(
np.arange(1, m1prodshape + 1, 1))
m2prodshape = (np.shape(m2)[0] * np.shape(m2)[1])
m2L = m2prodshape * (m2.flatten() - 1) + np.uint32(
np.arange(1, m2prodshape + 1, 1))
# Calculation broken up for clarity
H1 = np.reshape(horz_data.flatten()[m1L - 1], np.shape(m1))
H2 = np.reshape(horz_data.flatten()[m2L - 1], np.shape(m2))
H = np.minimum(H1, H2)
# Calculate cast shadows
# In ModImLam the original strict formulatuion:
# "bc[H < solar_zen] = 1"
# was relaxed to compute a bit larger, following Sirguey et al. 2009
# but it overestimates the cast shadows for the Alps
bc = np.ones(shape=np.shape(H)) # Initialise
sza_deg = np.rad2deg(sza)
bc[H < sza_deg] = 0
# bc[H < sza_deg + (-0.406 * sza_deg)] = 1
# Use a morphological operation (erode) to clean shadow mask by removing
# scattered pixels
bc_fat = morphology.grey_dilation(bc, size=(3, 3))
bc = morphology.grey_erosion(bc_fat, size=(3, 3))
# Calculate the combination of cast and self as binary product
b = np.logical_and(bs, bc).astype(int)
return (b, bs, bc)
def contourcuts(image,maxdist=15,minrange=10,mincdist=20,sigma=1.0,debug=0,r=8,s=0.5):
if debug:
figure(1); clf(); imshow(image)
# start by computing the contours
contours = image2contours(image!=0)
# generate a mask for grayscale morphology
mask = s*ones((r,r))
mask[2:-2,2:-2] = 0
cuts = []
# now handle each (external) contour individually
for k,cs in enumerate(contours):
# compute a matrix of all the pairwise distances of pixels
# around the contour, then smooth it a little
ds = distance.cdist(cs,cs)
ds = filters.gaussian_filter(ds,(sigma,sigma),mode='wrap')
# compute a circulant matrix telling us the pathlength
# between any two pixels on the contour
n = len(cs)
l = abs(arange(n)-n/2.0)
l = l[0]-l
cds = linalg.circulant(l)
# find true local minima (exclude ridges) by using the
# structuring element above
ge = morphology.grey_erosion(ds,structure=mask,mode='wrap')
locs = (ds<=ge)
# restrict it to pairs of points that are closer than maxdist
locs *= (ds<maxdist)
# restrict it to paris of points that are separated by
# at least mincdist on the contour
locs *= (cds>=mincdist)
# label the remaining minima and locate them
locs,n = measurements.label(locs)
cms = measurements.center_of_mass(locs,locs,range(1,n+1))
# keep only on of each pair (in canonical ordering)
cms = [(int(i+0.5),int(j+0.5)) for i,j in cms if i<j]
for i,j in cms:
x0,y0 = cs[i]
x1,y1 = cs[j]
# keep only the near vertical ones
if abs(y1-y0)>abs(x1-x0):
color = 'r'
cuts.append((cs[i],cs[j]))
else:
color = 'b'
if debug:
print (x0,y0),(x1,y1)
figure(1); plot([x0,x1],[y0,y1],color)
if debug:
figure(2); clf(); ion(); imshow(locs!=0)
figure(3); clf(); imshow(minimum(ds,maxdist*1.5),interpolation='nearest')
ginput(1,0.1)
print "hit ENTER"; raw_input()
# now construct a cut image
cutimage = zeros(image.shape)
for ((x0,y0),(x1,y1)) in cuts:
image_draw_line(cutimage,y0,x0,y1,x1)
cutimage = filters.maximum_filter(cutimage,(3,3))
if debug:
figure(4); clf(); imshow(maximum(0,image-0.5*cutimage))
return cutimage
def gray_erosion(self, *args, **kw):
'''see scipy.ndimage.morphology.grey_erosion'''
return Image(_morphology.grey_erosion(self, *args, **kw)).convert_type(self.dtype)
def calculate_metrics(self):
assert self.contains_mahalanobis_distances, "Can't calculate ROC without mahalanobis distances calculated"
# (Name, ROC_AUC, AUC_PR, f1)
results = list()
extractor = os.path.basename(self.filename).replace(".h5", "")
gauss_filters = [
None, (0, 1, 1), (0, 2, 2), (1, 0, 0), (1, 1, 1), (1, 2, 2)
]
# gauss_filters = [None, (0,1,1), (0,2,2), (0,3,3),
# (1,0,0), (1,1,1), (1,2,2), (1,3,3),
# (2,0,0), (2,1,1), (2,2,2), (2,3,3)]
other_filters = [None, "erosion", "dilation"]
for metric_name, metric in self.METRICS.items():
relevant = metric.get_relevant(self)
labels = metric.get_labels(relevant)
for other_filter in other_filters:
for gauss_filter in gauss_filters:
# Don't compute gauss filters (in image space) for per sum metrics (they take the average anyways)
if metric_name.endswith(
"(sum)"
) and gauss_filter != None and gauss_filter[1] > 0:
continue
title = "Metrics for %s (%s, filter:%s + %s)" % (
extractor, metric_name, gauss_filter, other_filter)
logger.info("Calculating %s" % title)
scores = dict()
for n in sorted(self.mahalanobis_distances.dtype.names):
name = n.replace("fake", "simple")
maha = relevant.mahalanobis_distances[n]
if gauss_filter is not None:
maha = utils.gaussian_filter(maha,
sigma=gauss_filter)
if other_filter is not None:
struct = generate_binary_structure(2, 1)
if struct.ndim == 2:
z = np.zeros_like(struct, dtype=np.bool)
struct = np.stack((z, struct, z))
if other_filter == "erosion":
maha = grey_erosion(maha, structure=struct)
elif other_filter == "dilation":
maha = grey_dilation(maha, structure=struct)
scores[name] = metric.get_values(maha)
filename = os.path.join(
consts.METRICS_PATH, "%s_%s_%s_%s.jpg" %
(extractor, metric_name, gauss_filter, other_filter))
result = self.calculate_roc(title, labels, scores,
filename)
for model, roc_auc, auc_pr, max_f1, fpr0, fpr1, fpr2, fpr3, fpr4, fpr5 in result:
results.append(
(extractor, metric_name, model, gauss_filter,
other_filter, roc_auc, auc_pr, max_f1, fpr0, fpr1,
fpr2, fpr3, fpr4, fpr5))
return results
if (exponent!= 1):
print "- Nonlinear Stretching..."
# Data=((Data/float(Data.max()/theta) )**exponent*255/phi).astype(int)
Data=((Data/float(Data.max()/theta) )**exponent*255/phi).astype(int)
if(save):
pl.save("Stretching", Data)
if(opening):
print "- Morphological Opening..."
Data=grey_opening(Data, structure=Cross)
if(save):
pl.save("Opening", Data)
if(erosion):
print "- Morphological Erosion..."
Data=grey_erosion(Data, structure=Cross)
if(save):
pl.save("Erosion", Data)
if(closing):
print "- Morphological Closing..."
Data=grey_closing(Data, structure=Cross)
if(save):
pl.save("Closing", Data)
# Remark: one could keep on with other transformations, other kernels and so on
# To do so, I would reccomend to use ipython, and eventually load the partial results
FinalStep=Data
if(view):
view_slice(FinalStep,SizeX/2)
def _adapt(self,
video_idx,
frame_idx,
last_mask,
get_posteriors_fn,
adapt_flag=0):
"""
adapt_flag (int): 0:do not adapt, 1:adapt with hard labels based on teacher,
2:adapt on hard labels from last mask
"""
# Perform Mask erosion to reduce effect of false positive
eroded_mask = grey_erosion(last_mask,
size=(self.erosion_size, self.erosion_size,
1))
# Compute distance transform
dt = distance_transform_edt(numpy.logical_not(eroded_mask))
# Adaptation target initialize
adaptation_target = numpy.zeros_like(last_mask)
adaptation_target[:] = VOID_LABEL
# Retrieve current probability map to adapt with
current_posteriors = get_posteriors_fn()
if adapt_flag == 2:
positives = current_posteriors[:, :,
1] > self.posterior_positive_threshold
elif adapt_flag == 1:
positives = last_mask == 1
if self.use_positives:
adaptation_target[positives] = 1
# Threshold based on distance transform
threshold = self.distance_negative_threshold
negatives = dt > threshold
if self.use_negatives:
adaptation_target[negatives] = 0
do_adaptation = eroded_mask.sum() > 0
# Save adaptation targets for debugging
if self.debug:
adaptation_target_visualization = adaptation_target.copy()
adaptation_target_visualization[adaptation_target == 1] = 128
if not do_adaptation:
adaptation_target_visualization[:] = VOID_LABEL
from scipy.misc import imsave
folder = self.val_data.video_tag().replace("__", "/")
imsave(
"forwarded/" + self.model + "/valid/" + folder +
"/adaptation_%05d.png" % frame_idx,
numpy.squeeze(adaptation_target_visualization))
self.train_data.set_video_idx(video_idx)
# Start Adapting based on number of adaptation_steps
for idx in xrange(self.n_adaptation_steps):
do_step = True
#if idx % self.adaptation_interval == 0:
if do_adaptation:
feed_dict = self.train_data.feed_dict_for_video_frame(
frame_idx, with_annotations=True)
feed_dict[self.train_data.get_label_placeholder(
)] = adaptation_target
loss_scale = self.adaptation_loss_scale
adaption_frame_idx = frame_idx
else:
do_step = False
if do_step:
loss, _, n_imgs = self.trainer.train_step(
epoch=idx,
feed_dict=feed_dict,
loss_scale=loss_scale,
learning_rate=self.adaptation_learning_rate)
assert n_imgs == 1
print >> log.v4, "adapting on frame", adaption_frame_idx, "of sequence", video_idx + 1, \
self.train_data.video_tag(video_idx), "loss:", loss
if do_adaptation:
return negatives
else:
return None
def contourcuts(image,maxdist=15,minrange=10,mincdist=20,sigma=1.0,debug=0,r=8,s=0.5):
if debug:
figure(1); clf(); imshow(image)
# start by computing the contours
contours = image2contours(image!=0)
# generate a mask for grayscale morphology
mask = s*ones((r,r))
mask[2:-2,2:-2] = 0
cuts = []
# now handle each (external) contour individually
for k,cs in enumerate(contours):
# compute a matrix of all the pairwise distances of pixels
# around the contour, then smooth it a little
ds = distance.cdist(cs,cs)
ds = filters.gaussian_filter(ds,(sigma,sigma),mode='wrap')
# compute a circulant matrix telling us the pathlength
# between any two pixels on the contour
n = len(cs)
l = abs(arange(n)-n/2.0)
l = l[0]-l
cds = linalg.circulant(l)
# find true local minima (exclude ridges) by using the
# structuring element above
ge = morphology.grey_erosion(ds,structure=mask,mode='wrap')
locs = (ds<=ge)
# restrict it to pairs of points that are closer than maxdist
locs *= (ds<maxdist)
# restrict it to paris of points that are separated by
# at least mincdist on the contour
locs *= (cds>=mincdist)
# label the remaining minima and locate them
locs,n = measurements.label(locs)
cms = measurements.center_of_mass(locs,locs,range(1,n+1))
# keep only on of each pair (in canonical ordering)
cms = [(int(i+0.5),int(j+0.5)) for i,j in cms if i<j]
for i,j in cms:
x0,y0 = cs[i]
x1,y1 = cs[j]
# keep only the near vertical ones
if abs(y1-y0)>abs(x1-x0):
color = 'r'
cuts.append((cs[i],cs[j]))
else:
color = 'b'
if debug:
print (x0,y0),(x1,y1)
figure(1); plot([x0,x1],[y0,y1],color)
if debug:
figure(2); clf(); ion(); imshow(locs!=0)
figure(3); clf(); imshow(minimum(ds,maxdist*1.5),interpolation='nearest')
ginput(1,0.1)
print "hit ENTER"; raw_input()
# now construct a cut image
cutimage = zeros(image.shape)
for ((x0,y0),(x1,y1)) in cuts:
image_draw_line(cutimage,y0,x0,y1,x1)
cutimage = filters.maximum_filter(cutimage,(3,3))
if debug:
figure(4); clf(); imshow(maximum(0,image-0.5*cutimage))
return cutimage
cv2.imwrite(cur_save_root + '00000.png', output)
for j in range(1, cur_seq_range):
img_file = osp.join(cur_seq_image_path, '%05d' % j + '.jpg')
cur_img, ori_img_size = load_image_label_davis17(
img_file, cfg['crop_size'])
cur_img = Variable(cur_img).cuda()
output = model.forward(cur_img, model_weights, train_mode, bn_params)
output = sigmoid(interp(output)).cpu().data[0, 0].numpy()
# use current frame to generate label for next frame
gen_label = np.zeros_like(output)
gen_label[:] = adapt_ignore_label
eroded_mask = grey_erosion(last_mask,
size=(adapt_erosion_size,
adapt_erosion_size))
eroded_mask[eroded_mask < 0.1] = 0
dt = distance_transform_edt(np.logical_not(eroded_mask))
gen_label[output > pos_pred_thres] = 1
gen_label[dt > neg_dist_thres] = 0
do_adapt = eroded_mask.sum() > 0
if adapt_debug:
if do_adapt:
gen_label_vis = gen_label.copy()
gen_label_vis[gen_label == 1] = 128
else:
gen_label_vis[:] = adapt_ignore_label
return keypoints, np.array(descrs)
img1 = cv2.imread("masked_points.jpg", 0)
img2 = cv2.imread("masked_disks.jpg", 0)
# pixel-wise mean taken over an image ensemble
mean1 = cv2.imread("masked_mean1.jpg", 0)
mean2 = cv2.imread("masked_mean2.jpg", 0)
# turn the points into disks
img1 -= mean1
img2 -= mean2
img1 = grey_erosion(img1, size=(3, 3))
# flip and scale the image
img1 = np.hflip(img1)
img1 = imresize(0.5)
# locate intensity peaks. wherever they are, black them out and replace them
with a large circle in a blank image
virtualdisks = np.zeros(img2.shape)
max_threshold = 200
virtualradius = 60
eraserradius = 13
while np.max(img1) > 200:
cy, cx = np.argmax(img1, axis=0), np.argmax(img1, axis=1)
rr, cc = circle(cy, cx, virtualradius)
def get_data_from_image_id(
image_id, coco_obj, img_size, base_path, grayscale=False,
generate_negative_points=True, aug_pipeline=None,
toggle_bin_mask=True,
toggle_joints=True,
toggle_dp_seg=True,
toggle_dp_points=True,
toggle_instance_offsets=True
):
all_results = {}
res = get_image_data(
coco_obj,
image_id,
base_path,
toggle_image=True,
toggle_bin_mask=toggle_bin_mask,
toggle_joints=toggle_joints,
toggle_dp_seg=toggle_dp_seg,
toggle_dp_points=toggle_dp_points,
toggle_instance_offsets=toggle_instance_offsets
)
# image
image = res['image']
xyhw_box = (0, 0, image.shape[0], image.shape[1]) # y, x, h, w
xyhw_box = bbu.extend_xyhw_to_ratio(xyhw_box, img_size[1]/img_size[0])
xyhw_box = bbu.round_bbox_params(xyhw_box)
padded_img = imu.pad_to_bbox(image, xyhw_box, mode='mean')
new_bbox = (max(0, xyhw_box[0]), max(0, xyhw_box[1]), xyhw_box[2], xyhw_box[3])
offset_x, offset_y = xyhw_box[0], xyhw_box[1]
padded_crop = padded_img[new_bbox[0]:new_bbox[0]+new_bbox[2], new_bbox[1]:new_bbox[1]+new_bbox[3]]
old_shape = padded_crop.shape[:]
resized_crop = cv2.resize(padded_crop, img_size[::-1], interpolation=cv2.INTER_AREA)
all_results['image'] = resized_crop
# keypoints
if toggle_joints:
all_keypoints = res['joints'].copy()
all_shifted_pts = all_keypoints - [offset_x, offset_y, 0, 0] # coordinates in the crops frame
all_shifted_pts = all_shifted_pts[ # remove ousiders
(0 <= all_shifted_pts[:, 0]) &
(all_shifted_pts[:, 0] < padded_crop.shape[0]) &
(0 <= all_shifted_pts[:, 1]) &
(all_shifted_pts[:, 1] < padded_crop.shape[1])
]
# rescale keypoints
all_rescaled_pts = all_shifted_pts * [img_size[0]/old_shape[0], img_size[1]/old_shape[1], 1., 1.]
all_results['joints'] = all_rescaled_pts
# fix joints loss region
joints_loss_region = res['joints_loss_region']
joints_loss_region = imu.pad_to_bbox(joints_loss_region, xyhw_box, mode='constant', cval=1) # pad with 1. to collect loss from no mask outside the image
joints_loss_region = joints_loss_region[new_bbox[0]:new_bbox[0]+new_bbox[2], new_bbox[1]:new_bbox[1]+new_bbox[3]]
joints_loss_region = cv2.resize(joints_loss_region.astype(np.uint8), img_size[::-1], interpolation=cv2.INTER_NEAREST).astype(np.float32)
all_results['joints_loss_region'] = grey_erosion(joints_loss_region, 5)
# pad, crop and resize bin_mask
if toggle_bin_mask:
_bin_mask = res['bin_mask']
_bin_mask = imu.pad_to_bbox(_bin_mask, xyhw_box, mode='constant')
_bin_mask = _bin_mask[new_bbox[0]:new_bbox[0]+new_bbox[2], new_bbox[1]:new_bbox[1]+new_bbox[3]]
bin_mask = cv2.resize(_bin_mask.astype(np.uint8), img_size[::-1], interpolation=cv2.INTER_NEAREST).astype(np.float32)
all_results['bin_mask'] = bin_mask
# pad, crop and resize dp_mask
if toggle_dp_seg:
_dp_mask = res['dp_mask']
_dp_mask = imu.pad_to_bbox(_dp_mask, xyhw_box, mode='constant')
_dp_mask[:, :, 0] = ~np.logical_or.reduce(_dp_mask[:, :, 1:], axis=2)
_dp_mask = _dp_mask[new_bbox[0]:new_bbox[0]+new_bbox[2], new_bbox[1]:new_bbox[1]+new_bbox[3]]
dp_mask = resize(_dp_mask.astype(np.float32), img_size, order=0, mode='edge', anti_aliasing=False).astype(np.float32)
# dp_mask = softmax(dp_mask, axis=2)
dp_mask /= dp_mask.sum(axis=2)[:, :, None]
all_results['dp_mask'] = dp_mask
# rescale densepose points
if toggle_dp_points:
_dp_points = res['dp_points'].copy()
_dp_points = _dp_points - [offset_x, offset_y, 0., 0., 0.] # coordinates in the crops frame
_dp_points = _dp_points[ # remove ousiders
(0 <= _dp_points[:, 0]) &
(_dp_points[:, 0] < padded_crop.shape[0]) &
(0 <= _dp_points[:, 1]) &
(_dp_points[:, 1] < padded_crop.shape[1])
]
# rescale keypoints
dp_points = _dp_points * [img_size[0]/old_shape[0], img_size[1]/old_shape[1], 1., 1., 1.]
if generate_negative_points:
dp_points = np.concatenate([dp_points, generate_neg_coords(bin_mask.astype(np.bool) | (~dp_mask[:, :, 0].astype(np.bool)))], axis=0)
all_results['dp_points'] = dp_points
# instance offsets
if toggle_instance_offsets:
inst_offsets, head_points = res['instance_offsets'], res['head_points']
inst_offsets = imu.pad_to_bbox(inst_offsets, xyhw_box, mode='constant')
inst_offsets = inst_offsets[new_bbox[0]:new_bbox[0]+new_bbox[2], new_bbox[1]:new_bbox[1]+new_bbox[3]]
# inst_offsets /= list(inst_offsets.shape[:2]) # normalize offsets
inst_offsets = cv2.resize(inst_offsets, img_size[::-1], interpolation=cv2.INTER_NEAREST)
all_results['instance_offsets'] = inst_offsets
if head_points != []:
head_points = head_points - [offset_x, offset_y]
head_points = head_points * [img_size[0]/old_shape[0], img_size[1]/old_shape[1]]
all_results['head_points'] = head_points
################## AUGMENTS HERE ###################
if aug_pipeline is not None:
all_results = aug_pipeline(all_results)
if toggle_instance_offsets:
pts = all_results['head_points']
all_results['instance_offsets_loss_region'] = (all_results['instance_offsets'] > 0).astype(np.float32)
offset_mask = np.zeros((img_size[0], img_size[1], 2), dtype=np.float32)
if pts != []:
for i in range(pts.shape[0]):
head_point = pts[i]
# convert instance maps into offset maps and loss region
xx, yy = np.where(all_results['instance_offsets'] == i + 1)
offset_mask[xx, yy, 0] = head_point[0] - xx
offset_mask[xx, yy, 1] = head_point[1] - yy
all_results['instance_offsets'] = offset_mask / [img_size[0], img_size[1]]
del all_results['head_points']
return all_results
def _adapt(self, video_idx, frame_idx, last_mask, get_posteriors_fn,
kframe_lst):
eroded_mask = grey_erosion(last_mask,
size=(self.erosion_size, self.erosion_size,
1))
dt = distance_transform_edt(numpy.logical_not(eroded_mask))
adaptation_target2 = last_mask
adaptation_target = numpy.zeros_like(last_mask)
adaptation_target[:] = VOID_LABEL
current_posteriors = get_posteriors_fn()
positives = current_posteriors[:, :,
1] > self.posterior_positive_threshold
if self.use_positives:
adaptation_target[positives] = 1
threshold = self.distance_negative_threshold
negatives = dt > threshold
if self.use_negatives:
adaptation_target[negatives] = 0
do_adaptation = eroded_mask.sum() > 0
if self.debug:
adaptation_target_visualization = adaptation_target.copy()
adaptation_target_visualization[adaptation_target == 1] = 128
if not do_adaptation:
adaptation_target_visualization[:] = VOID_LABEL
from scipy.misc import imsave
folder = self.val_data.video_tag().replace("__", "/")
imsave(
"forwarded/" + self.model + "/valid/" + folder +
"/adaptation_%05d.png" % frame_idx,
numpy.squeeze(adaptation_target_visualization))
self.train_data.set_video_idx(video_idx)
threshold_ = 0.020
#print "self.n_adaptation_steps",n_adaptation_steps
for idx in xrange(self.n_adaptation_steps): #
print idx
do_step = True
if idx % self.adaptation_interval == 0: #adaptation_interval
if do_adaptation:
print("NewIter")
#--------------------pre-frame-result-----------------
feed_dict = self.train_data.feed_dict_for_video_frame(
frame_idx, with_annotations=True)
feed_dict[self.train_data.get_label_placeholder(
)] = adaptation_target #
loss_scale = self.adaptation_loss_scale * 5
adaption_frame_idx = frame_idx
else:
print >> log.v4, "skipping current frame adaptation, since the target seems to be lost"
do_step = False
elif idx % self.adaptation_interval == 1:
print "idx % self.adaptation_interval =1"
if frame_idx >= 1:
print "-----------"
key_list = self.extractBorderFrame(video_idx,
kframe_lst[-1],
frame_idx - 1,
threshold_)
key = key_list[-1]
if key not in kframe_lst:
kframe_lst.append(key)
feed_dict = self.train_data.feed_dict_for_video_frame(
kframe_lst[-1], with_annotations=True)
loss_scale = self.adaptation_loss_scale
adaption_frame_idx = kframe_lst[-1]
elif idx % self.adaptation_interval == 2:
print "idx % self.adaptation_interval=2"
if len(kframe_lst) > 2:
feed_dict = self.train_data.feed_dict_for_video_frame(
kframe_lst[-2], with_annotations=True)
adaption_frame_idx = kframe_lst[-2]
if kframe_lst[-2] - kframe_lst[-1] > 5:
loss_scale = self.adaptation_loss_scale * 100
else:
loss_scale = self.adaptation_loss_scale
elif idx % self.adaptation_interval == 3:
print "idx % self.adaptation_interval=3"
if len(kframe_lst) > 3:
feed_dict = self.train_data.feed_dict_for_video_frame(
kframe_lst[-3], with_annotations=True)
adaption_frame_idx = kframe_lst[-3]
if kframe_lst[-3] - kframe_lst[-1] > 10:
loss_scale = self.adaptation_loss_scale * 100
else:
loss_scale = self.adaptation_loss_scale
else:
print "else"
# mix in first frame to avoid drift
# (do this even if we think the target is lost, since then this can help to find back the target)
feed_dict = self.train_data.feed_dict_for_video_frame(
0, with_annotations=True)
loss_scale = 1.0
adaption_frame_idx = 0
if do_step:
loss, measures, n_imgs = self.trainer.train_step(
epoch=idx,
feed_dict=feed_dict,
loss_scale=loss_scale,
learning_rate=self.adaptation_learning_rate)
iou = Measures.calc_iou(measures, n_imgs, [0])
assert n_imgs == 1
print >> log.v4, "adapting on frame", adaption_frame_idx, "of sequence", video_idx + 1, \
self.train_data.video_tag(video_idx), "loss:", loss,"iou:",iou
if do_adaptation:
return negatives, kframe_lst
else:
return None, kframe_lst
def __maha__(self, x=None, only_refresh_image=False):
image = np.zeros((350, 480, 3), dtype=np.uint8)
if self.model_index > 0 and self.patches.contains_mahalanobis_distances:
font = cv2.FONT_HERSHEY_SIMPLEX
fontScale = 0.5
thickness = 1
model = sorted(self.patches.mahalanobis_distances.dtype.names)[self.model_index - 1]
cv2.putText(image,"Model:", (10, 20), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
cv2.putText(image, model, (65, 20), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
cv2.putText(image,"Filter:", (10, 50), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
if not only_refresh_image:
self.patches.mahalanobis_distances_filtered[:] = self.patches.mahalanobis_distances[model]
sigma_0 = (cv2.getTrackbarPos("0_gaussian_0", self.WINDOWS_MAHA),
cv2.getTrackbarPos("0_gaussian_1", self.WINDOWS_MAHA),
cv2.getTrackbarPos("0_gaussian_2", self.WINDOWS_MAHA))
if sigma_0 != (0, 0, 0):
cv2.putText(image, "gaussian (%i, %i, %i)" % sigma_0, (65, 50), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
if not only_refresh_image:
self.patches.mahalanobis_distances_filtered = utils.gaussian_filter(self.patches.mahalanobis_distances_filtered, sigma=sigma_0)
erosion_dilation = cv2.getTrackbarPos("1_erosion_dilation", self.WINDOWS_MAHA)
if erosion_dilation > 0:
struct = generate_binary_structure(cv2.getTrackbarPos("1_erosion_dilation_structure_rank", self.WINDOWS_MAHA),
cv2.getTrackbarPos("1_erosion_dilation_structure_connectivity", self.WINDOWS_MAHA))
if struct.ndim == 2:
z = np.zeros_like(struct, dtype=np.bool)
struct = np.stack((z, struct, z))
if erosion_dilation == 1:
cv2.putText(image, "erosion", (65, 80), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
if not only_refresh_image:
self.patches.mahalanobis_distances_filtered = grey_erosion(self.patches.mahalanobis_distances_filtered, structure=struct)
elif erosion_dilation == 2:
cv2.putText(image, "dilation", (65, 80), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
if not only_refresh_image:
self.patches.mahalanobis_distances_filtered = grey_dilation(self.patches.mahalanobis_distances_filtered, structure=struct)
for (z, x, y) in np.ndindex(struct.shape):
cv2.putText(image, str(int(struct[z, x, y])), (150 + y * 15 + z * 60, 80 + x * 15), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
sigma_2 = (cv2.getTrackbarPos("2_gaussian_0", self.WINDOWS_MAHA),
cv2.getTrackbarPos("2_gaussian_1", self.WINDOWS_MAHA),
cv2.getTrackbarPos("2_gaussian_2", self.WINDOWS_MAHA))
if sigma_2 != (0, 0, 0):
cv2.putText(image, "gaussian (%i, %i, %i)" % sigma_2, (65, 140), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
if not only_refresh_image:
self.patches.mahalanobis_distances_filtered = utils.gaussian_filter(self.patches.mahalanobis_distances_filtered, sigma=sigma_2)
# Add some statistics
threshold = float(cv2.getTrackbarPos("threshold", self.WINDOWS_MAHA)) / 10000.0
cv2.putText(image, " TPR FPR Threshold", (10, 190), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
self._metrics_ax1.clear()
# self._metrics_ax2.clear()
self._metrics_ax1.set_yscale("log")
for i, (metric_name, metric) in enumerate(sorted(PatchArray.METRICS.items())):
labels = metric.get_labels(self.patches)
scores = metric.get_values(self.patches.mahalanobis_distances_filtered)
if metric.current_threshold == -1:
m = np.max(scores)
metric.current_threshold = m
else:
m = metric.current_threshold
thresh = m * threshold
negavites = scores[labels == 1]
positives = scores[labels == 2]
false_negavites = np.count_nonzero(negavites >= thresh)
true_positives = np.count_nonzero(positives >= thresh)
tpr = true_positives / float(positives.size) * 100.0 if float(positives.size) > 0 else 0
fpr = false_negavites / float(negavites.size) * 100.0 if float(negavites.size) > 0 else 0
if metric_name != "per patch" and i == cv2.getTrackbarPos("metric", self.WINDOWS_MAHA):
self._labels = labels
self._scores = scores
self._thresh = thresh
for r in np.reshape(np.diff(np.r_[0, labels == 0, 0]).nonzero()[0], (-1,2)):
self._metrics_ax1.axvspan(r[0], r[1], facecolor='black', alpha=0.1)
for r in np.reshape(np.diff(np.r_[0, np.logical_and(labels == 2, scores >= thresh), 0]).nonzero()[0], (-1,2)):
self._metrics_ax1.axvspan(r[0], r[1], facecolor='g', alpha=0.2)
for r in np.reshape(np.diff(np.r_[0, np.logical_and(labels == 1, scores >= thresh), 0]).nonzero()[0], (-1,2)):
self._metrics_ax1.axvspan(r[0], r[1], facecolor='r', alpha=0.2)
for r in np.reshape(np.diff(np.r_[0, np.logical_and(labels == 2, scores < thresh), 0]).nonzero()[0], (-1,2)):
self._metrics_ax1.axvspan(r[0], r[1], facecolor='b', alpha=0.2)
self._metrics_ax1.set_ylim(0, np.max(scores))
self._metrics_ax1.plot(scores, lw=1, label=metric_name, color="black")
# self._metrics_ax1.axvline(x=self.index, linewidth=0.5, color="black")
self._metrics_ax1.axhline(y=thresh, linewidth=0.5, color="black")
self._metrics_fig.suptitle(metric_name)
if not only_refresh_image:
self._histogram_ax1.clear()
self._histogram_ax2.clear()
# r = (np.nanmin(self.patches.mahalanobis_distances_filtered), np.nanmax(self.patches.mahalanobis_distances_filtered))
self._histogram_ax1.set_title("No anomaly")
self._histogram_ax2.set_title("Anomaly")
self._histogram_fig.suptitle("Mahalanobis distances")
_, bins, _ = self._histogram_ax1.hist(negavites.ravel(), bins=200)
self._histogram_ax2.hist(positives.ravel(), bins=bins)
self._histogram_fig.canvas.draw()
cv2.putText(image, metric_name, (40, 220 + i*30), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
cv2.putText(image, "%.2f" % tpr, (200, 220 + i*30), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
cv2.putText(image, "%.2f" % fpr, (300, 220 + i*30), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
cv2.putText(image, "%.2f" % thresh, (400, 220 + i*30), font, fontScale, (255,255,255), thickness, lineType=cv2.LINE_AA)
self._metrics_fig.canvas.draw()
self.__draw__()
cv2.imshow(self.WINDOWS_MAHA, image)
