def view_raw(args):
data = read_raw(args.view_count)
for name, _, cls, bbox in data:
print(name, cls, bbox)
for d in group(data, 36):
# display raw images
_, images, _, _ = zip(*d)
show_image(merge_images_with_border([cv2.resize(i, (150, 150)) for i in images]))
def view_train(args):
_, conf = get_model_and_config(args)
train_data, _ = get_train_validation_raw(conf.validation_raw_examples_ratio)
train_data = batch_generator(train_data, conf.batch_size, use_hog=args.view_hog)
viewed = 0
for images, _ in train_data:
show_image(merge_images_with_border(images))
viewed += conf.batch_size
if viewed >= args.view_count:
break
def test_model(trained_ops, test_set_sz, learning_rate, batch_size, latent_dim,
tau, inf_layers, gen_layers, restore_path):
test_init = trained_ops[0]
ops = trained_ops[1]
next_batch = trained_ops[2]
saver = tf.train.Saver()
with tf.Session() as sess:
mean_loss = 0
mean_rec = 0
mean_kl = 0
test_iters = int(test_set_sz / batch_size)
lat = None
lab = None
gen_imgs = None
rec_imgs = None
saver.restore(sess, './saved_models/model.ckpt')
sess.run(test_init)
print("\nTesting...")
for i in range(test_iters):
loss, rec, kl, z, x_hat, x_gen, y = sess.run(
(ops[0], ops[1], ops[2], ops[4], ops[5], ops[6],
next_batch[1]))
mean_loss += loss / test_iters
mean_rec += rec / test_iters
mean_kl += kl / test_iters
if lat is None:
lat = z[0]
lab = y
else:
lat = np.append(lat, z[0], axis=0)
lab = np.append(lab, y, axis=0)
# reconstruct samples and generate new ones
if i == test_iters - 3 or i == test_iters - 2 or i == test_iters - 1:
if gen_imgs is None:
rec_imgs = x_hat.reshape((batch_size * 28, 28))
gen_imgs = x_gen.reshape((batch_size * 28, 28))
else:
rec_imgs = np.append(rec_imgs,
x_hat.reshape((batch_size * 28, 28)),
axis=1)
gen_imgs = np.append(gen_imgs,
x_gen.reshape((batch_size * 28, 28)),
axis=1)
show_image(gen_imgs, './images/generated.png')
show_image(rec_imgs, './images/reconstructed.png')
print("\tloss: {:.3f}".format(mean_loss),
"\tmean reconstruction: {:.3f}".format(mean_rec),
"\tmean KL: {:.3f}".format(mean_kl))
plot_pca(10, lat, lab, './figures/test_latent_space')
def view_test(args):
_, conf = get_model_and_config(args)
_, raw_test_data = get_train_validation_raw(conf.validation_raw_examples_ratio)
images, _ = get_unaugmented(raw_test_data, use_hog=args.view_hog)
features, labels = get_unaugmented(raw_test_data, use_hog=conf.use_hog)
if args.eval:
if not conf.model_dir or not args.model_name:
raise ValueError("Reguired model-dir and model-name for testset evaluation.")
prediction = predict_classes(args, features, labels)
is_correct = [np.argmax(lbl) == np.argmax(pred) for lbl, pred in zip(labels, prediction)]
assert len(is_correct) == len(images)
print('good answers:', sum(is_correct))
print('accuracy:', sum(is_correct) / len(is_correct))
images = decorate_images(images, is_correct)
show_image(merge_images(images, scale=0.2))
else:
# show_image(merge_images_with_border([cv2.resize(i, (100, 100)) for i in images]))
show_image(merge_images_with_border(images))
def check_moving_detection():
import moving_detection
rgb_im = image_processing.load_image("falling balls and cylinder",
"rgb_" + str(0) + ".png")
depth_im = image_processing.load_image("falling balls and cylinder",
"depth_" + str(0) + ".png", "depth")
start = time.time()
mog = moving_detection.Fast_RGBD_MoG(rgb_im,
depth_im,
number_of_gaussians=5)
print("initialization: ", time.time() - start)
for i in range(4):
rgb_im = image_processing.load_image("falling balls and cylinder",
"rgb_" + str(i + 1) + ".png")
depth_im = image_processing.load_image("falling balls and cylinder",
"depth_" + str(i + 1) + ".png",
"depth")
start = time.time()
mask = mog.get_mask(rgb_im, depth_im)
print("frame updating: ", time.time() - start)
visualization.show_image(mask / 255)
BATCH_SIZE = 10
test_data = raw_file_idx3_process("data/test/t10k-images-idx3-ubyte")
test_label = raw_file_idx1_process("data/test/t10k-labels-idx1-ubyte")
test_set = Torch_data_set(test_data, test_label)
test_loader = tud.DataLoader(test_set, batch_size=BATCH_SIZE, shuffle=False)
total = 0
bingo = 0
print("\n")
with torch.no_grad():
for index, data in enumerate(test_loader):
inputs, label = data
inputs, label = inputs.to(device), label.to(device)
outputs = net(inputs)
_, predicted = torch.max(outputs.data, 1)
# 对total和bingo进行修改
total += label.size(0)
bingo += (predicted == label).sum().item()
# 显示第128号的图像
if index == 117:
print(predicted.cpu().numpy())
inputs = [inp.reshape(28, 28) for inp in inputs.cpu()]
v.show_image(inputs, 2, 5)
# 打印正确率
print("\rAccuracy: {:.2f}".format(bingo / total), end='')
title_fs = 10
zone_color = 'darkgray'
zone_lw = 1.3
line_color = zone_color
line_lw = zone_lw
park_color = 'orange'
train_color = 'orangered'
test_color = 'deepskyblue'
fig = plt.figure(figsize=(13, 7))
gs = fig.add_gridspec(2, 3, hspace=0.25, wspace=0.0)
# whole image
ax_whole = fig.add_subplot(gs[:, 0])
ax_whole.set_title('Taï National Park', fontsize=title_fs)
show_image(img_whole, meta_whole['transform'], ax=ax_whole, band_idx=[2, 1, 0])
for p in polygons_Tai:
ax_whole.add_patch(
matplotlib.patches.Polygon(
p, **{
'linewidth': 1,
'facecolor': (0, 0, 0, 0),
'edgecolor': 'orange'
}))
ax_whole.add_patch(
matplotlib.patches.Rectangle(nw_bounds[0:2],
width=np.abs(nw_bounds[0] - nw_bounds[2]),
height=np.abs(nw_bounds[1] - nw_bounds[3]),
linewidth=zone_lw,
facecolor=(0, 0, 0, 0),
edgecolor=zone_color))
def initialize_expectation_maximization(filepath_1, filepath_2, components, iterations, subtraction_threshold):
"""
perform initialization step
:param filepath_1: path to first image (required)
:param filepath_2: path to second image (optional)
:param components: number of components
:param iterations: number of iterations
:return: means: list of initial mean values
:return: variances: list of initial variance values
:return: stdevs: list of initial stdev values
:return: weights: list of initial weight values
:return: log_likelihoods: list of initial log_likelihood values
"""
# initialize data_matrix based on number of images provided
image_1 = np.divide(np.float64(cv2.imread(filepath_1)), np.float64(255)) # read image from disk. normalize.
data_matrix = image_1[:, :, 0] # use only first matrix layer in case user provides non-grayscale input.
if filepath_2 is not None:
# given 2 images as input, data matrix is their CIE94 difference
image_2 = np.divide(np.float64(cv2.imread(filepath_2)), np.float64(255)) # read image from disk. normalize.
cie94_difference = \
image_processing.cie94_distance_metric(image_processing.xyz_to_lab(image_processing.lrgb_to_xyz(image_1)),
image_processing.xyz_to_lab(image_processing.lrgb_to_xyz(image_2)))
data_matrix = (cie94_difference + np.min(cie94_difference)) / (np.max(cie94_difference) - np.min(cie94_difference)) # normalize
# initialize algorithm state based on parameters
if subtraction_threshold is not None and components == 2: # BG-FG thresholded subtraction initialization
# initialization based on subtraction threshold (only appropriate for 2 images input and 2 components)
print("Initializing GMM state based on thresholded subtraction.")
# visualization.show_image((1, 1, 1), "CIE94 Difference Image", cie76_difference, vmin=np.min(cie76_difference),
# vmax=np.max(cie94_difference), postprocessing=False)
cie94_segmentation_bg = np.int32(cie94_difference <= subtraction_threshold)
cie94_segmentation_fg = np.int32(cie94_difference > subtraction_threshold)
visualization.show_image((1, 1, 1), "CIE94 Segmentation w/ Threshold="+str(subtraction_threshold), cie94_segmentation_fg,
vmin=np.min(cie94_segmentation_fg),
vmax=np.max(cie94_segmentation_fg), postprocessing=False)
# compute initial algorithm state
rows = data_matrix.shape[0]
cols = data_matrix.shape[1]
fg_num_pixels = np.count_nonzero(cie94_segmentation_fg)
bg_num_pixels = np.count_nonzero(cie94_segmentation_bg)
fg_mean = np.sum(cie94_segmentation_fg * data_matrix) / fg_num_pixels
bg_mean = np.sum(cie94_segmentation_bg * data_matrix) / bg_num_pixels
fg_segmentation_array = np.reshape(cie94_segmentation_fg, (rows*cols))
bg_segmentation_array = np.reshape(cie94_segmentation_bg, (rows*cols))
data_matrix_array = np.reshape(data_matrix, (rows*cols))
fg_pixels_array = [data_matrix_array[i] for i in range(0, data_matrix_array.shape[0]) if fg_segmentation_array[i]]
bg_pixels_array = [data_matrix_array[i] for i in range(0, data_matrix_array.shape[0]) if bg_segmentation_array[i]]
init_means = [bg_mean, fg_mean]
init_variances = [np.var(bg_pixels_array), np.var(fg_pixels_array)]
init_stdevs = np.sqrt(init_variances)
init_weights = [np.float64(bg_num_pixels) / np.float64(bg_num_pixels + fg_num_pixels),
np.float64(fg_num_pixels) / np.float64(bg_num_pixels + fg_num_pixels)]
else: # arbitrary initialization based on even component spacing assumption
print("Initializing GMM state based on evenly spaced component assumption.")
init_means = np.linspace(0, 1, components) # assume initial component means are evenly spaced in pixel value domain
init_variance = np.float64(10 / 255) # initialized as explained in GMM tutorial paper
init_variances = np.float64(np.ones(components)) * init_variance # initial variance is 10/255 - quite small
init_stdevs = np.sqrt(init_variances)
init_weights = np.ones(components)
# log likelihoods do not have the concept of initialization; simply create a list to be populated by EM algorithm
init_log_likelihoods = np.zeros(iterations)
print("Completed Initialization")
print("init_means=", str(init_means),
"\ninit_variances=", str(init_variances),
"\ninit_stdevs=", str(init_stdevs),
"\ninit_weights=", str(init_weights))
return data_matrix, init_means, init_variances, init_stdevs, init_weights, init_log_likelihoods
def view_aug(args):
data = read_raw(2)[1:2]
images, _ = next(iter(batch_generator(data, args.view_count, use_hog=args.view_hog)))
show_image(get_unaugmented(data, use_hog=args.view_hog)[0][0])
show_image(merge_images_with_border(images))
# ---------------------------------------------------------------------------------------------------------------
# %% Generate new features
img_train = preprocessing(np.stack(band, axis=2))
# %% vizualize new features
Nband_kept = 10 - 1
fig, axs = plt.subplots(4,
4,
figsize=(12, 10),
gridspec_kw={
'hspace': 0.1,
'wspace': 0.1
})
show_image(img_train[:, :, Nband_kept + 1],
meta_train['transform'],
ax=axs[0][0],
cmap='PiYG')
axs[0][0].set_title('NDVI', fontsize=10)
axs[0][0].set_ylabel('Northing [m]')
axs[0][0].set_xticklabels([])
show_image(img_train[:, :, Nband_kept + 2],
meta_train['transform'],
ax=axs[0][1],
cmap='PiYG')
axs[0][1].set_title('green NDVI', fontsize=10)
axs[0][1].set_xticklabels([])
axs[0][1].set_yticklabels([])
titles_part1 = [
'entropy: ', 'Opening reconstruction: ', 'closing reconstruction: ',
'LBP: '
# Visualize results
fig = plt.figure(figsize=(12, 16))
gs = plt.GridSpec(4, 3, wspace=0.2, hspace=0.2)
title_fs = 11
park_color = 'orange'
pred_fc = 'crimson'
pred_ec = None
pred_alpha = 1
offset_w = [9, 9, 9, 9, 9, 9, 9, 9, 9, 9]
offset_h = [5, 5, 5, 5, 5, 9, 5, 5, 5, 9]
# overview img
ax_overview = fig.add_subplot(gs[0:2, 0])
ax_overview.set_title('Taï National Park', fontsize=title_fs)
show_image(img_whole,
meta_whole['transform'],
ax=ax_overview,
band_idx=[2, 1, 0])
for p in polygons_park:
ax_overview.add_patch(
matplotlib.patches.Polygon(
p, **{
'linewidth': 1,
'facecolor': (0, 0, 0, 0),
'edgecolor': 'orange'
}))
for i, bound in enumerate(bounds):
ax_overview.add_patch(matplotlib.patches.Rectangle(bound[0:2], \
width=np.abs(bound[0] - bound[2]), \
height=np.abs(bound[1] - bound[3]), \
linewidth=1, facecolor=(0,0,0,0), edgecolor='darkgray'))
ax_overview.text(bound[2] - offset_w[i] * w,
********HOW TO USE***********
pip install -r requirements.txt
0 - Open config.py
1 - Set R, C and FEATURE_3D
2 - Set IMAGE_NAME and booleans RESIZE and BLUR (for preprocessing)
3 - Run
"""
from config import *
from utils import pre_process, save_image, concatenate_images
from imageSegmentation import segmIm
from visualization import cv2, show_image
if __name__ == '__main__':
start_time = time.time()
img, img_origin = pre_process()
segmented_img, num_peaks = segmIm(img, R, C, FEATURE_3D)
final_im = concatenate_images(
[cv2.cvtColor(img_origin, cv2.COLOR_RGB2BGR), img, segmented_img])
seconds = time.time() - start_time
print("--- %s seconds ---" % seconds)
if SAVE:
SAVE_NAME = IMAGE_NAME + '_r' + str(R) + '_c' + str(C) + dim_name(
) + blur_name() + '_' + str(round(seconds)) + 'sec.jpg'
save_image(final_im, SAVE_NAME)
else:
show_image(final_im)
def show_hausdorff_distance(shape_distance_id):
bitmap = render_hausdorff_distance(shape_distance_id)
visualization.show_image(bitmap)