def gen_layers_for_image(i, img):
"""
Generate laplacian pyramids and normalize every channel of every
pyramid.
"""
img = resize(img[:, :, :], requested_shape)
new_imgs = yuv_laplacian_norm(img, requested_shape, 3)
return i, new_imgs
def gen_layers_for_image_hog(i, img):
"""
Generate laplacian pyramids and normalize every channel of every
pyramid of RGB.
Calc HOG of depth at every scale.
"""
img = resize(img[:, :, :], requested_shape)
rgb_img = img[:, :, 0:3]
depth_img = img[:, :, 3]
# transform
rgb_imgs = yuv_laplacian_norm(rgb_img, requested_shape, n_layers=3)
# depth_img = calc_hog(depth_img)
depth_img = depth_img.astype('float32') / 255.0
new_imgs = []
for img in rgb_imgs:
shp = (img.shape[1], img.shape[2])
new_img = np.concatenate(
(img, resize(depth_img, shp).reshape((1, shp[0], shp[1]))), axis=0)
new_imgs.append(new_img)
return i, new_imgs
def get_stats(results, y, n_classes, dont_care_classes,
fnames_path, dataset_type, show=False, log=True,
postproc=None, postproc_params=None):
assert(len(results) == y.shape[0])
fnames = []
with open(fnames_path, 'r') as f:
fnames = f.read().split('\n')
care_classes = np.ones((n_classes), dtype='int8')
care_classes[dont_care_classes] = 0
correct = np.zeros((n_classes), dtype='int32')
total = np.zeros((n_classes), dtype='int32')
for ind, img in enumerate(results):
curr_y = y[ind]
img_up = zoom(img, 4, order=0)
img_old = img_up
assert(img_up.shape[0] == SHAPE[0])
images_path = IMGS
orig_img = open_image(images_path, fnames[ind] + '.jpg')
orig_img = resize(orig_img, SHAPE)
# apply postprocessing
if postproc is not None:
img_up = postproc(orig_img, img_up, **postproc_params)
if show and ind < IMGS_TO_SHOW:
bounds = np.linspace(0, 30, 31)
# cmap = plt.get_cmap('Paired')
cmap = plt.get_cmap('gist_ncar')
norm = matplotlib.colors.BoundaryNorm(bounds, cmap.N)
plt.subplot(2, 2, 1)
plt.axis('off')
plt.imshow(orig_img)
plt.subplot(2, 2, 2)
plt.axis('off')
plt.imshow(curr_y, cmap=cmap, norm=norm)
plt.subplot(2, 2, 3)
plt.axis('off')
plt.imshow(img_old, cmap=cmap, norm=norm)
plt.subplot(2, 2, 4)
plt.axis('off')
plt.imshow(img_up, cmap=cmap, norm=norm)
plt.show()
plt.draw()
# count correct pixels
for i in range(n_classes):
if np.any(curr_y == i):
correct[i] += np.sum(np.equal(img_up[curr_y == i],
curr_y[curr_y == i]))
total[i] += np.sum(curr_y == i)
care_correct = correct * care_classes
care_total = total * care_classes
class_accuracy = calc_class_accuracy(care_correct, care_total)
total_acc = (np.sum(care_correct, dtype='float32') /
np.sum(care_total)) * 100.
if log:
logger.info('total pixel accuracy %f %%', total_acc)
logger.info('mean class accuracy: %f %%',
class_accuracy * 100.)
logger.info('per class accuracies: %s',
correct.astype('float32') / total)
return total_acc
def mark_image(i, img, cc, requested_shape):
logger.info("Marking image %d", i)
img = resize(img, requested_shape, inter=0)
assert(img.shape[:2] == requested_shape)
layers = process_out(img, cc, requested_shape)
return i, layers
def get_stats(results,
y,
n_classes,
dont_care_classes,
fnames_path,
dataset_type,
show=False,
log=True,
postproc=None,
postproc_params=None):
assert (len(results) == y.shape[0])
fnames = []
with open(fnames_path, 'r') as f:
fnames = f.read().split('\n')
care_classes = np.ones((n_classes), dtype='int8')
care_classes[dont_care_classes] = 0
correct = np.zeros((n_classes), dtype='int32')
total = np.zeros((n_classes), dtype='int32')
for ind, img in enumerate(results):
curr_y = y[ind]
img_up = zoom(img, 4, order=0)
img_old = img_up
assert (img_up.shape[0] == SHAPE[0])
images_path = IMGS
orig_img = open_image(images_path, fnames[ind] + '.jpg')
orig_img = resize(orig_img, SHAPE)
# apply postprocessing
if postproc is not None:
img_up = postproc(orig_img, img_up, **postproc_params)
if show and ind < IMGS_TO_SHOW:
bounds = np.linspace(0, 30, 31)
# cmap = plt.get_cmap('Paired')
cmap = plt.get_cmap('gist_ncar')
norm = matplotlib.colors.BoundaryNorm(bounds, cmap.N)
plt.subplot(2, 2, 1)
plt.axis('off')
plt.imshow(orig_img)
plt.subplot(2, 2, 2)
plt.axis('off')
plt.imshow(curr_y, cmap=cmap, norm=norm)
plt.subplot(2, 2, 3)
plt.axis('off')
plt.imshow(img_old, cmap=cmap, norm=norm)
plt.subplot(2, 2, 4)
plt.axis('off')
plt.imshow(img_up, cmap=cmap, norm=norm)
plt.show()
plt.draw()
# count correct pixels
for i in range(n_classes):
if np.any(curr_y == i):
correct[i] += np.sum(
np.equal(img_up[curr_y == i], curr_y[curr_y == i]))
total[i] += np.sum(curr_y == i)
care_correct = correct * care_classes
care_total = total * care_classes
class_accuracy = calc_class_accuracy(care_correct, care_total)
total_acc = (np.sum(care_correct, dtype='float32') /
np.sum(care_total)) * 100.
if log:
logger.info('total pixel accuracy %f %%', total_acc)
logger.info('mean class accuracy: %f %%', class_accuracy * 100.)
logger.info('per class accuracies: %s',
correct.astype('float32') / total)
return total_acc