def bwdist(
img,
method=cv2.DIST_L2,
dist_mask=cv2.DIST_MASK_5,
label_type=cv2.DIST_LABEL_CCOMP,
ravel=True,
):
"""Mimics Matlab's bwdist function, similar to OpenCV's distanceTransform()
but with different output.
https://www.mathworks.com/help/images/ref/bwdist.html
Available metrics:
https://docs.opencv.org/3.4/d7/d1b/group__imgproc__misc.html#gaa2bfbebbc5c320526897996aafa1d8eb
Available distance masks:
https://docs.opencv.org/3.4/d7/d1b/group__imgproc__misc.html#gaaa68392323ccf7fad87570e41259b497
Available label types:
https://docs.opencv.org/3.4/d7/d1b/group__imgproc__misc.html#ga3fe343d63844c40318ee627bd1c1c42f
"""
flip = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY_INV)[1]
dist, labeled = cv2.distanceTransformWithLabels(flip, method, dist_mask)
if ravel: # return linear indices if ravel == True (default)
idx = np.zeros(img.shape, dtype=np.intp)
idx_func = np.flatnonzero
else: # return two-channel indices if ravel == False
idx = np.zeros((*img.shape, 2), dtype=np.intp)
idx_func = lambda masked: np.dstack(np.where(masked))
for l in np.unique(labeled):
mask = labeled == l
idx[mask] = idx_func(img * mask)
return dist, idx
def colorate_template_map_by_segm(template, segm, scale):
texture_image = np.zeros((2048, 2048))
map_height = np.int32(texture_image.shape[0] / scale)
map_width = np.int32(texture_image.shape[1] / scale)
face_indices_map = np.zeros((map_height, map_width, 3))
text_coords = template.vt[:, :2]
text_coords[:, 0] *= map_width
text_coords[:, 1] = (1 - text_coords[:, 1]) * map_height
text_coords = np.int32(text_coords)
parts = range(1 + np.max(segm))
for part in parts:
indices = np.where(segm == part)[0]
color = np.int32(np.random.rand(3) * 255)
while np.all(color == [0, 0, 0]):
color = np.int32(np.random.rand(3) * 255)
for face in indices:
points = text_coords[template.ft[face]]
points = np.int32([points]).squeeze()
cv2.fillConvexPoly(face_indices_map, points, color.tolist())
b_coords_map = face_indices_map
unmapped = (np.sum((b_coords_map.reshape(
b_coords_map.shape[0] * b_coords_map.shape[1], 3) == -np.zeros(3)),
axis=1) == 3).reshape(b_coords_map.shape[0],
b_coords_map.shape[1])
dst, lab = cv2.distanceTransformWithLabels(unmapped.astype(np.uint8),
distanceType=1,
maskSize=3,
labelType=1)
full_texture = face_indices_map[np.logical_not(unmapped)][lab - 1]
return full_texture
def watershed_segmentation(rgb_img, mask, distance=10):
"""Uses the watershed algorithm to detect boundary of objects. Needs a marker file which specifies area which is
object (white), background (grey), unknown area (black).
Inputs:
rgb_img = image to perform watershed on needs to be 3D (i.e. np.shape = x,y,z not np.shape = x,y)
mask = binary image, single channel, object in white and background black
distance = min_distance of local maximum
Returns:
analysis_images = list of output images
:param rgb_img: numpy.ndarray
:param mask: numpy.ndarray
:param distance: int
:return analysis_images: list
"""
params.device += 1
# Store debug mode
debug = params.debug
params.debug = None
dist_transform = cv2.distanceTransformWithLabels(mask, cv2.DIST_L2, maskSize=0)[0]
localMax = peak_local_max(dist_transform, indices=False, min_distance=distance, labels=mask)
markers = ndi.label(localMax, structure=np.ones((3, 3)))[0]
dist_transform1 = -dist_transform
labels = watershed(dist_transform1, markers, mask=mask)
img1 = np.copy(rgb_img)
for x in np.unique(labels):
rand_color = color_palette(len(np.unique(labels)))
img1[labels == x] = rand_color[x]
img2 = apply_mask(img1, mask, 'black')
joined = np.concatenate((img2, rgb_img), axis=1)
estimated_object_count = len(np.unique(markers)) - 1
# Reset debug mode
params.debug = debug
if params.debug == 'print':
print_image(dist_transform, os.path.join(params.debug_outdir, str(params.device) + '_watershed_dist_img.png'))
print_image(joined, os.path.join(params.debug_outdir, str(params.device) + '_watershed_img.png'))
elif params.debug == 'plot':
plot_image(dist_transform, cmap='gray')
plot_image(joined)
outputs.add_observation(variable='estimated_object_count', trait='estimated object count',
method='plantcv.plantcv.watershed', scale='none', datatype=int,
value=estimated_object_count, label='none')
# Store images
outputs.images.append([dist_transform, joined])
return joined
def compute_sdf(mask, signed=True, with_indices=True):
""" Returns a signed distance transform for a mask, and label indices.
:param mask: Zero values are exterior, non-zero values are interior area
"""
# Get interior mask by erosion, compute contour from that and invert for 0-level set
contour, interior = compute_contour(mask)
zero_level = (1 - contour).astype(np.uint8)
# Build mapping between 1D contour and label index
mask2idx = [-1]
mask2idx.extend(np.flatnonzero(contour))
if with_indices:
dt, idx = cv2.distanceTransformWithLabels(
zero_level,
cv2.DIST_L2,
cv2.DIST_MASK_PRECISE,
labelType=cv2.DIST_LABEL_PIXEL)
else:
dt = cv2.distanceTransform(zero_level, cv2.DIST_L2,
cv2.DIST_MASK_PRECISE)
if signed:
dt[interior.astype(bool)] *= -1 # Make interior negative
if not with_indices:
return dt
# Map back from pixel label to 1D image position
for r in range(idx.shape[0]):
for c in range(idx.shape[1]):
idx[r, c] = mask2idx[idx[r, c]]
return dt, idx, contour
def watershed_segmentation(rgb_img, mask, distance=10):
"""Uses the watershed algorithm to detect boundary of objects. Needs a marker file which specifies area which is
object (white), background (grey), unknown area (black).
Inputs:
rgb_img = image to perform watershed on needs to be 3D (i.e. np.shape = x,y,z not np.shape = x,y)
mask = binary image, single channel, object in white and background black
distance = min_distance of local maximum
Returns:
analysis_images = list of output images
:param rgb_img: numpy.ndarray
:param mask: numpy.ndarray
:param distance: int
:return analysis_images: list
"""
params.device += 1
# # Will be depricating opencv version 2
# if cv2.__version__[0] == '2':
# dist_transform = cv2.distanceTransform(mask, cv2.cv.CV_DIST_L2, maskSize=0)
# else:
dist_transform = cv2.distanceTransformWithLabels(mask, cv2.DIST_L2, maskSize=0)[0]
localMax = peak_local_max(dist_transform, indices=False, min_distance=distance, labels=mask)
markers = ndi.label(localMax, structure=np.ones((3, 3)))[0]
dist_transform1 = -dist_transform
labels = watershed(dist_transform1, markers, mask=mask)
img1 = np.copy(rgb_img)
for x in np.unique(labels):
rand_color = color_palette(len(np.unique(labels)))
img1[labels == x] = rand_color[x]
img2 = apply_mask(img1, mask, 'black')
joined = np.concatenate((img2, rgb_img), axis=1)
estimated_object_count = len(np.unique(markers)) - 1
analysis_image = []
analysis_image.append(joined)
if params.debug == 'print':
print_image(dist_transform, os.path.join(params.debug_outdir, str(params.device) + '_watershed_dist_img.png'))
print_image(joined, os.path.join(params.debug_outdir, str(params.device) + '_watershed_img.png'))
elif params.debug == 'plot':
plot_image(dist_transform, cmap='gray')
plot_image(joined)
outputs.add_observation(variable='estimated_object_count', trait='estimated object count',
method='plantcv.plantcv.watershed', scale='none', datatype=int,
value=estimated_object_count, label='none')
# Store images
outputs.images.append(analysis_image)
return analysis_image
def getRadiusCenterCircle(raw_dist):
dist_transform, label = cv2.distanceTransformWithLabels(raw_dist,
cv2.DIST_L2,
maskSize=5)
_, radius, _, center = cv2.minMaxLoc(dist_transform)
points = None
# 넘버링 여러 곳에 하는 기능 개발 중, 주석처리하였음
# dist_transform = cv2.distanceTransform(raw_dist, cv2.DIST_L2, maskSize=5)
# points = [list(dist_transform)[i] for i in range(0, len(dist_transform), 300) if list(dist_transform)[i] > 50]
# print(type(list(dist_transform)[50]), list(dist_transform)[50])
# ret, dist1 = cv2.threshold(dist_transform, 0.6*dist_transform.max(), 255, 0)
# print(f'dist_transform : {dist_transform}')
# print(f'label: {label}')
# points = []
# points = np.where(dist_transform > 10)
# for idx in range(0, len(dist_transform)-30, 30):
# _, radius, _, center = cv2.minMaxLoc(dist_transform[idx:idx+30])
# if radius > 10:
# points.append((radius, center))
# print(np.unique(np.where(label > 10)))
# result = cv2.normalize(dist_transform, None, 255, 0, cv2.NORM_MINMAX, cv2.CV_8UC1)
# minVal, maxVal, a, center = cv2.minMaxLoc(result)
return radius, center, points
def nearest_point(refined_lidar):
value_mask = np.asarray(1.0 - np.squeeze(refined_lidar) > 0.1).astype(
np.uint8)
dt, lbl = cv2.distanceTransformWithLabels(value_mask,
cv2.DIST_L1,
5,
labelType=cv2.DIST_LABEL_PIXEL)
return dt, lbl
def loadsympascal(self, imgidx, gtidx):
# load image and skeleton
image = cv2.imread(
'{}/SymPASCAL-by-KZ/{}'.format(self.data_dir, imgidx), 1)
skeleton = cv2.imread(
'{}/SymPASCAL-by-KZ/{}'.format(self.data_dir, gtidx), 0)
skeleton = (skeleton > 0).astype(np.uint8)
# normalization
image = image.astype(np.float32)
image -= self.mean
image = image.transpose((2, 0, 1))
# compute flux and dilmask
kernel = np.ones((15, 15), np.uint8)
dilmask = cv2.dilate(skeleton, kernel)
rev = 1 - skeleton
height = rev.shape[0]
width = rev.shape[1]
rev = (rev > 0).astype(np.uint8)
dst, labels = cv2.distanceTransformWithLabels(
rev,
cv2.cv.CV_DIST_L2,
cv2.cv.CV_DIST_MASK_PRECISE,
labelType=cv2.DIST_LABEL_PIXEL)
index = np.copy(labels)
index[rev > 0] = 0
place = np.argwhere(index > 0)
nearCord = place[labels - 1, :]
x = nearCord[:, :, 0]
y = nearCord[:, :, 1]
nearPixel = np.zeros((2, height, width))
nearPixel[0, :, :] = x
nearPixel[1, :, :] = y
grid = np.indices(rev.shape)
grid = grid.astype(float)
diff = grid - nearPixel
dist = np.sqrt(np.sum(diff**2, axis=0))
direction = np.zeros((2, height, width), dtype=np.float32)
direction[0, rev > 0] = np.divide(diff[0, rev > 0], dist[rev > 0])
direction[1, rev > 0] = np.divide(diff[1, rev > 0], dist[rev > 0])
direction[0] = direction[0] * (dilmask > 0)
direction[1] = direction[1] * (dilmask > 0)
flux = -1 * np.stack((direction[0], direction[1]))
dilmask = (dilmask > 0).astype(np.float32)
dilmask = dilmask[np.newaxis, ...]
return image, flux, dilmask
def disappear_instance(labels_source,
instances,
instance_ids=None,
only_objects=True,
swap_fraction=0.5,
**_):
print("Removing instances")
if instance_ids is None:
instance_uniq = np.unique(instances)
if only_objects:
instance_uniq_objects = instance_uniq[instance_uniq >= 24000]
else:
instance_uniq_objects = instance_uniq[instance_uniq >= 1]
if instance_uniq_objects.__len__() == 0:
return dict(labels_swapErr=labels_source.copy(), )
instance_ids = np.random.choice(instance_uniq_objects,
int(instance_uniq_objects.__len__() *
swap_fraction),
replace=False)
disappear_mask = np.any([instances == i for i in instance_ids], axis=0)
obj_classes = np.unique(labels_source[disappear_mask])
forbidden_class = remove_trainIDs
mask_forbidden_class = np.any(
[labels_source == j for j in forbidden_class], axis=0)
mask_no_label = mask_forbidden_class | disappear_mask
mask_label = np.logical_not(mask_no_label)
closest_dst, closest_labels = cv2.distanceTransformWithLabels(
mask_no_label.astype(np.uint8),
distanceType=cv2.DIST_L2,
maskSize=5,
labelType=cv2.DIST_LABEL_PIXEL,
)
BG_indices = closest_labels[mask_label]
BG_labels = labels_source[mask_label]
label_translation = np.zeros(labels_source.shape,
dtype=np.uint8).reshape(-1)
label_translation[BG_indices] = BG_labels
label_recon = labels_source.copy()
label_recon[disappear_mask] = label_translation[closest_labels.reshape(
labels_source.shape)[disappear_mask]]
result = dict(labels_swapErr=label_recon, )
return result
def update(dummy=None):
global need_update
need_update = False
thrs = cv2.getTrackbarPos('threshold', 'distrans')
mark = cv2.Canny(img, thrs, 3*thrs)
dist, labels = cv2.distanceTransformWithLabels(~mark, cv.CV_DIST_L2, 5)
if voronoi:
vis = cm[np.uint8(labels)]
else:
vis = cm[np.uint8(dist*2)]
vis[mark != 0] = 255
cv2.imshow('distrans', vis)
def update(dummy=None):
global need_update
need_update = False
thrs = cv2.getTrackbarPos('threshold', 'distrans')
mark = cv2.Canny(img, thrs, 3 * thrs)
dist, labels = cv2.distanceTransformWithLabels(~mark, cv2.DIST_L2, 5)
if voronoi:
vis = cm[np.uint8(labels)]
else:
vis = cm[np.uint8(dist * 2)]
vis[mark != 0] = 255
cv2.imshow('distrans', vis)
def spread_labels(labels, maxdist=9999999):
"""Spread the given labels to the background."""
#distances,features = morphology.distance_transform_edt(labels==0,return_distances=1,return_indices=1)
#indexes = features[0]*labels.shape[1]+features[1]
#spread = labels.ravel()[indexes.ravel()].reshape(*labels.shape)
distances, indexes = cv2.distanceTransformWithLabels(
array(labels == 0, uint8),
cv2.DIST_L2,
cv2.DIST_MASK_PRECISE,
labelType=cv2.DIST_LABEL_PIXEL)
spread = labels[where(labels > 0)][indexes - 1]
spread *= (distances < maxdist)
return spread
def nearest_filling(depth_map):
zero_mask = depth_map < 1e-3
dist, labels = cv2.distanceTransformWithLabels(
src=zero_mask.astype(np.uint8),
distanceType=cv2.DIST_L2,
maskSize=cv2.DIST_MASK_3,
labelType=cv2.DIST_LABEL_PIXEL)
far_mask = dist > 35.
labels = labels - 1
depth_vec = depth_map[~zero_mask]
intered_depth = depth_vec[labels]
intered_depth[far_mask] = 0.
return intered_depth
def run(controller):
maps_initialized = False
count = 0
while (True):
#sleepで更新速度を制御
time.sleep(1)
frame = controller.frame()
image = frame.images[0]
if image.is_valid:
if not maps_initialized:
left_coordinates, left_coefficients = convert_distortion_maps(
frame.images[0])
right_coordinates, right_coefficients = convert_distortion_maps(
frame.images[1])
maps_initialized = True
undistorted_left = undistort(image, left_coordinates,
left_coefficients, 400, 400)
undistorted_right = undistort(image, right_coordinates,
right_coefficients, 400, 400)
#画像を２値化（白黒に処理）
ret, hand = cv2.threshold(undistorted_right, 70, 255,
cv2.THRESH_BINARY)
my_hand = hand[80:320, 40:360]
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - my_hand, cv2.DIST_L2, 5)
img_label = np.uint8(img_label) & my_hand
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最も広い白領域のみを残す
this_hand = np.uint8(img_label == m) * 255
# 得られた二値化画像を画面に表示
cv2.imshow('hand', this_hand)
#cv2.imshow('hand', undistorted_right)
cv2.imwrite("ImageData/5/img_five{0:03d}.png".format(count),
this_hand)
count += 1
if cv2.waitKey(1) & 0xFF == ord('q'):
break
def run(controller):
maps_initialized = False
while(True):
#sleepで更新速度を制御
time.sleep(0.1)
frame = controller.frame()
image = frame.images[0]
if image.is_valid:
if not maps_initialized:
left_coordinates, left_coefficients = convert_distortion_maps(frame.images[0])
right_coordinates, right_coefficients = convert_distortion_maps(frame.images[1])
maps_initialized = True
undistorted_left = undistort(image, left_coordinates, left_coefficients, 400, 400)
undistorted_right = undistort(image, right_coordinates, right_coefficients, 400, 400)
#画像を２値化（白黒に処理）
ret,hand = cv2.threshold(undistorted_right,80,255,cv2.THRESH_BINARY)
my_hand = hand[80:320,40:360]
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
img_dist, img_label = cv2.distanceTransformWithLabels(255-my_hand, cv2.DIST_L2, 5)
img_label = np.uint8(img_label) & my_hand
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最も広い白領域のみを残す
this_hand = np.uint8(img_label == m)*255
# 最大の白領域からscikit-learnに入力するためのベクトルを取得
hand_vector = getImageVector(this_hand)
# 学習済のニューラルネットワークから分類結果を取得
result = clf.predict(hand_vector)
# 分類結果を表示
print(janken_class[result[0]])
# 得られた二値化画像を画面に表示
cv2.imshow('hand', this_hand)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
def label2flux(label):
"""
Create direction_field of corresponding label
Returns:
direction_field: shape (2, h, w), weight_matrix: (h, w)
"""
label = label.copy()
label = label.astype(np.int32)
height, width = label.shape
label += 1
categories = np.unique(label)
if 0 in categories:
raise ValueError('invalid category')
label = cv2.copyMakeBorder(label, 1, 1, 1, 1, cv2.BORDER_CONSTANT, value=0)
weight_matrix = np.zeros((height+2, width+2), dtype=np.float32)
direction_field = np.zeros((2, height+2, width+2), dtype=np.float32)
for category in categories:
img = (label == category).astype(np.uint8)
weight_matrix[img > 0] = 1. / np.sqrt(img.sum())
_, labels = cv2.distanceTransformWithLabels(img, cv2.DIST_L2, cv2.DIST_MASK_PRECISE, labelType=cv2.DIST_LABEL_PIXEL)
index = np.copy(labels)
index[img > 0] = 0
place = np.argwhere(index > 0)
nearCord = place[labels-1,:]
x = nearCord[:, :, 0]
y = nearCord[:, :, 1]
nearPixel = np.zeros((2, height+2, width+2))
nearPixel[0,:,:] = x
nearPixel[1,:,:] = y
grid = np.indices(img.shape)
grid = grid.astype(float)
diff = grid - nearPixel
direction_field[:, img > 0] = diff[:, img > 0]
weight_matrix = weight_matrix[1:-1, 1:-1]
direction_field = direction_field[:, 1:-1, 1:-1]
return direction_field, weight_matrix
def Distance_Transform_simple(lidar):
# only use the nearest neighbor to fill in the depth map
lidar = np.squeeze(lidar)
height, width = np.shape(lidar)
mask_lidar = lidar > 0.9
value_mask = np.asarray(1.0 - np.squeeze(mask_lidar)).astype(np.uint8)
dt, lbl = cv2.distanceTransformWithLabels(value_mask,
cv2.DIST_L1,
5,
labelType=cv2.DIST_LABEL_PIXEL)
with_value = mask_lidar
depth_list = np.squeeze(lidar[with_value])
label_list = np.reshape(lbl, [1, height * width])
depth_list_all = depth_list[label_list - 1]
depth_map = np.reshape(depth_list_all, (height, width))
return depth_map
def fill_spherical(range_image):
# fill in spherical image for calculating normal vector
height, width = np.shape(range_image)[:2]
value_mask = np.asarray(1.0 - np.squeeze(range_image) > 0.1).astype(
np.uint8)
dt, lbl = cv2.distanceTransformWithLabels(value_mask,
cv2.DIST_L1,
5,
labelType=cv2.DIST_LABEL_PIXEL)
with_value = np.squeeze(range_image) > 0.1
depth_list = np.squeeze(range_image)[with_value]
label_list = np.reshape(lbl, [1, height * width])
depth_list_all = depth_list[label_list - 1]
depth_map = np.reshape(depth_list_all, (height, width))
depth_map = cv2.GaussianBlur(depth_map, (7, 7), 0)
#depth_map=range_image*with_value+depth_map*(1-with_value)
return depth_map
def watershedSegmentation(wb, dataCube, plantMaterial, lighting, destination,
datacubeName):
"""
Segements the plant material with watershed segmentation algorithm,
saves the segmented data in CSV files.
:param wb: (Spectronon workbench)
:param dataCube: (numpy array) BIL Datacube
:param plantMaterial: (numpy array) Locations of the plant material
:param lighting: (string) Illumination description string
:param destination: (string) Directory in which to save the data
:param datacubeName: (string) Name of the current data cube
:return: (None)
"""
lines, samples, bands = dataCube.shape
plantMaterialGray = plantMaterial
plantMaterial = cv2.merge(
(plantMaterialGray, plantMaterialGray, plantMaterialGray))
# Remove noise from the image
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(plantMaterialGray, cv2.MORPH_OPEN, kernel)
# Locate the for-sure background area
sure_bg = cv2.dilate(opening, kernel)
# Finding the for-sure foreground area
dist_transform, labels = cv2.distanceTransformWithLabels(
opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.4 * dist_transform.max(),
255, 0)
sure_fg = cv2.morphologyEx(sure_fg, cv2.MORPH_CLOSE, kernel)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero. We will have the segmented "leaves"
# present in the image
markers[unknown == 255] = 0
markers = cv2.watershed(plantMaterial, markers)
numMarkers = max(markers.flatten())
plt.imshow(markers)
plt.show()
saveData = wb.postQuestion("Continue with saving the segmented data?",
title="Segmentation Okay?")
if not saveData:
return
# Save the data for each leaf in a separate csv file (skip the background)
for leaf in xrange(2, numMarkers):
outputData = dataCube[markers == leaf, :]
if len(outputData.flatten()) / 240 < 2:
break
filename = datacubeName + '_' + lighting + '_' + str(leaf).zfill(
4) + ".csv"
fileLocation = os.path.join(destination, filename)
np.savetxt(fileLocation, outputData, delimiter=',')
def main(N):
savefile = "result.pkl"
# 手の認識用パラメータ（HチャンネルとSチャンネルとを二値化するための条件）
hmin = 0
hmax = 30 # 15-40程度にセット
smin = 50
janken_class = ['グー', 'チョキ', 'パー']
class_count = [0, 0, 0]
#savefile = sys.argv[1]
try:
clf = joblib.load(savefile)
except IOError:
print('学習済ファイル{0}を開けません'.format(savefile))
sys.exit()
with picamera.PiCamera() as camera:
with picamera.array.PiRGBArray(camera) as stream:
# カメラの解像度を320x240にセット
camera.resolution = (320, 240)
# カメラのフレームレートを15fpsにセット
camera.framerate = 15
# ホワイトバランスをfluorescent(蛍光灯)モードにセット
camera.awb_mode = 'fluorescent'
print('認識を開始します')
for i in range(N * 2):
sleep(0.5)
# stream.arrayにBGRの順で映像データを格納
camera.capture(stream, 'bgr', use_video_port=True)
# 映像データをHSV形式に変換
hsv = cv2.cvtColor(stream.array, cv2.COLOR_BGR2HSV)
# HSV形式からHチャンネルとSチャンネルの画像を得る
hsv_channels = cv2.split(hsv)
h_channel = hsv_channels[0]
s_channel = hsv_channels[1]
# Hチャンネルを平滑化
h_binary = cv2.GaussianBlur(h_channel, (5, 5), 0)
# Hチャンネルの二値化画像を作成
# hmin～hmaxの範囲を255（白）に、それ以外を0（黒）に
ret, h_binary = cv2.threshold(h_binary, hmax, 255,
cv2.THRESH_TOZERO_INV)
ret, h_binary = cv2.threshold(h_binary, hmin, 255,
cv2.THRESH_BINARY)
# Sチャンネルの二値化画像を作成
# smin～255の範囲を255（白）に、それ以外を0に（黒）に
ret, s_binary = cv2.threshold(s_channel, smin, 255,
cv2.THRESH_BINARY)
# HチャンネルとSチャンネルの二値化画像のANDをとる
# HチャンネルとSチャンネルの両方で255（白）の領域のみ白となる
hs_and = h_binary & s_binary
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - hs_and, cv2.cv.CV_DIST_L2, 5)
img_label = np.uint8(img_label) & hs_and
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最も広い白領域のみを残す
hand = np.uint8(img_label == m) * 255
# 最大の白領域からscikit-learnに入力するためのベクトルを取得
hand_vector = getImageVector(hand)
# 学習済のニューラルネットワークから分類結果を取得
result = clf.predict(hand_vector)
class_count[result[0]] += 1
# 分類結果を表示
print(janken_class[result[0]])
# 得られた二値化画像を画面に表示
#cv2.imshow('hand', hand)
# 'q'を入力でアプリケーション終了
#key = cv2.waitKey(1)
#if key & 0xFF == ord('q'):
# break
# streamをリセット
stream.seek(0)
stream.truncate()
cv2.destroyAllWindows()
#一番多かった手をリターン
return class_count.index(max(class_count))
ret, h_binary = cv2.threshold(h_binary, hmax, 255,
cv2.THRESH_TOZERO_INV)
ret, h_binary = cv2.threshold(h_binary, hmin, 255,
cv2.THRESH_BINARY)
# Sチャンネルの二値化画像を作成
# smin～255の範囲を255（白）に、それ以外を0に（黒）に
ret, s_binary = cv2.threshold(s_channel, smin, 255,
cv2.THRESH_BINARY)
# HチャンネルとSチャンネルの二値化画像のANDをとる
# HチャンネルとSチャンネルの両方で255（白）の領域のみ白となる
hs_and = h_binary & s_binary
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - hs_and, cv2.cv.CV_DIST_L2, 5)
img_label = np.uint8(img_label) & hs_and
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最も広い白領域のみを残す
hand = np.uint8(img_label == m) * 255
# 表示して動作チェックするため h_channel, s_channel, h_binary, s_binary を結合
hs = np.concatenate((h_channel, h_binary), axis=0)
hs_bin = np.concatenate((s_channel, s_binary), axis=0)
hs_final = np.concatenate((hs_and, hand), axis=0)
def watershed_segmentation(device, img, mask, distance=10, filename=False, debug=None):
"""Uses the watershed algorithm to detect boundary of objects. Needs a marker file which specifies area which is
object (white), background (grey), unknown area (black).
Inputs:
device = device number. Used to count steps in the pipeline
img = image to perform watershed on needs to be 3D (i.e. np.shape = x,y,z not np.shape = x,y)
mask = binary image, single channel, object in white and background black
distance = min_distance of local maximum
filename = if user wants to output analysis images change filenames from false
debug = None, print, or plot. Print = save to file, Plot = print to screen.
Returns:
device = device number
watershed_header = shape data table headers
watershed_data = shape data table values
analysis_images = list of output images
:param device: int
:param img: numpy array
:param mask: numpy array
:param distance: int
:param filename: str
:param debug: str
:return device: int
:return watershed_header: list
:return watershed_data: list
:return analysis_images: list
"""
if cv2.__version__[0] == '2':
dist_transform = cv2.distanceTransform(mask, cv2.cv.CV_DIST_L2, maskSize=0)
else:
dist_transform = cv2.distanceTransformWithLabels(mask, cv2.DIST_L2, maskSize=0)[0]
localMax = peak_local_max(dist_transform, indices=False, min_distance=distance, labels=mask)
markers = ndi.label(localMax, structure=np.ones((3, 3)))[0]
dist_transform1 = -dist_transform
labels = watershed(dist_transform1, markers, mask=mask)
img1 = np.copy(img)
for x in np.unique(labels):
rand_color = color_palette(len(np.unique(labels)))
img1[labels == x] = rand_color[x]
device, img2 = apply_mask(img1, mask, 'black', device, debug=None)
joined = np.concatenate((img2, img), axis=1)
estimated_object_count = len(np.unique(markers)) - 1
analysis_images = []
if filename != False:
out_file = str(filename[0:-4]) + '_watershed.jpg'
print_image(joined, out_file)
analysis_images.append(['IMAGE', 'watershed', out_file])
watershed_header = (
'HEADER_WATERSHED',
'estimated_object_count'
)
watershed_data = (
'WATERSHED_DATA',
estimated_object_count
)
if debug == 'print':
print_image(dist_transform, str(device) + '_watershed_dist_img.png')
print_image(joined, str(device) + '_watershed_img.png')
elif debug == 'plot':
plot_image(dist_transform, cmap='gray')
plot_image(joined)
return device, watershed_header, watershed_data, analysis_images
marker2.color.g = 0.0
marker2.color.b = 1.0
marker2.pose.orientation.w = 1.0
marker2.lifetime = rospy.Duration()
#########################################
print "precomputing distance field .."
t1 = time.time()
#outsource to cython
obstaclemap = util.get_obstaclemap(mapdata)
(distance_field,
labels) = cv2.distanceTransformWithLabels(obstaclemap,
cv2.DIST_L2,
cv2.DIST_MASK_PRECISE,
labelType=cv2.DIST_LABEL_PIXEL)
distance_field = np.array(distance_field, dtype=float)
t2 = time.time()
print "computed distance field in ", (t2 - t1)
pf = util.ParticleFilter(mapdata, distance_field, 15000)
with open(logfile) as f:
i = 0
for line in f:
message = line.split(" ")[:-1] #last element is timestamp
plt.title("Histogram_flat")
plt.plot(hist_flat)
plt.xlabel("# Bin")
plt.ylabel("% pixel")
plt.xlim([0, 512])
chans = cv2.split(image)
colors = ("b", "g", "r")
plt.figure(2)
plt.title("'Flattened' Color Histogram")
plt.xlabel("Bins")
plt.ylabel("# of Pixels")
features = []
# loop over the image channels
for (chan, color) in zip(chans, colors):
# create a histogram for the current channel and
# concatenate the resulting histograms for each
# channel
hist = cv2.calcHist([chan], [0], None, [256], [0, 256])
features.extend(hist)
# plot the histogram
plt.plot(hist, color=color)
plt.xlim([0, 256])
print(np.array(features).flatten().shape[0])
plt.show()
cv2.waitKey()
cv2.destroyAllWindows()
cv2.distanceTransformWithLabels()
def white_black():
from PIL import Image
import cv2
import numpy as np
from scipy import stats
hmin = 0
hmax = 30 # 15-40程度にセット
smin = 50
while True:
img_src = cv2.imread("img_tsumami000.jpg", 1) #テストの場合はここに画像へのパス
# 映像データをHSV形式に変換
hsv_src = cv2.cvtColor(img_src, cv2.COLOR_BGR2HSV)
# HSV形式からHチャンネルとSチャンネルの画像を得る
hsv_channels = cv2.split(hsv_src)
h_channel = hsv_channels[0]
s_channel = hsv_channels[1]
# Hチャンネルを平滑化
h_binary = cv2.GaussianBlur(h_channel, (5, 5), 0)
# Hチャンネルの二値化画像を作成
# hmin～hmaxの範囲を255（白）に、それ以外を0（黒）に
ret, h_binary = cv2.threshold(h_binary, hmax, 255,
cv2.THRESH_TOZERO_INV)
ret, h_binary = cv2.threshold(h_binary, hmin, 255, cv2.THRESH_BINARY)
# Sチャンネルの二値化画像を作成
# smin～255の範囲を255（白）に、それ以外を0に（黒）に
ret, s_binary = cv2.threshold(s_channel, smin, 255, cv2.THRESH_BINARY)
# HチャンネルとSチャンネルの二値化画像のANDをとる
# HチャンネルとSチャンネルの両方で255（白）の領域のみ白となる
hs_and = h_binary & s_binary
path = "after2.jpg"
cv2.imwrite(path, hs_and)
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - hs_and, cv2.DIST_L2, 5)
img_label = np.uint8(img_label) & hs_and
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最も広い白領域のみを残す
hand = np.uint8(img_label == m) * 255
key = cv2.waitKey(1) & 0xFF
# 得られた二値化画像を画面に表示
cv2.imshow("hand", hand)
cv2.moveWindow('hand', 800, 200)
cv2.imwrite("img_tsumami_test.jpg", hand)
if key == ord('q'):
break
cv2.destroyAllWindows()
return hand
def watershed_segmentation(rgb_img, mask, distance=10):
"""Uses the watershed algorithm to detect boundary of objects. Needs a marker file which specifies area which is
object (white), background (grey), unknown area (black).
Inputs:
rgb_img = image to perform watershed on needs to be 3D (i.e. np.shape = x,y,z not np.shape = x,y)
mask = binary image, single channel, object in white and background black
distance = min_distance of local maximum
Returns:
watershed_header = shape data table headers
watershed_data = shape data table values
analysis_images = list of output images
:param rgb_img: numpy.ndarray
:param mask: numpy.ndarray
:param distance: int
:return watershed_header: list
:return watershed_data: list
:return analysis_images: list
"""
params.device += 1
# # Will be depricating opencv version 2
# if cv2.__version__[0] == '2':
# dist_transform = cv2.distanceTransform(mask, cv2.cv.CV_DIST_L2, maskSize=0)
# else:
dist_transform = cv2.distanceTransformWithLabels(mask, cv2.DIST_L2, maskSize=0)[0]
localMax = peak_local_max(dist_transform, indices=False, min_distance=distance, labels=mask)
markers = ndi.label(localMax, structure=np.ones((3, 3)))[0]
dist_transform1 = -dist_transform
labels = watershed(dist_transform1, markers, mask=mask)
img1 = np.copy(rgb_img)
for x in np.unique(labels):
rand_color = color_palette(len(np.unique(labels)))
img1[labels == x] = rand_color[x]
img2 = apply_mask(img1, mask, 'black')
joined = np.concatenate((img2, rgb_img), axis=1)
estimated_object_count = len(np.unique(markers)) - 1
analysis_image = []
analysis_image.append(joined)
watershed_header = (
'HEADER_WATERSHED',
'estimated_object_count'
)
watershed_data = (
'WATERSHED_DATA',
estimated_object_count
)
if params.debug == 'print':
print_image(dist_transform, os.path.join(params.debug_outdir, str(params.device) + '_watershed_dist_img.png'))
print_image(joined, os.path.join(params.debug_outdir, str(params.device) + '_watershed_img.png'))
elif params.debug == 'plot':
plot_image(dist_transform, cmap='gray')
plot_image(joined)
# Store into global measurements
if not 'watershed' in outputs.measurements:
outputs.measurements['watershed'] = {}
outputs.measurements['watershed']['estimated_object_count'] = estimated_object_count
# Store images
outputs.images.append(analysis_image)
return watershed_header, watershed_data, analysis_image
def model_fn(model,
data,
criterion,
perfermance=False,
vis=False,
device="cuda",
epoch=0,
num_class=4,
w=None):
# imgs, gts, _ = data
imgs, gts = data[:2]
imgs = imgs.to(device)
gts = torch.squeeze(gts, 1).to(device)
gts = gts.to(device)
loss = 0
tb_dict = {}
disp_dict = {}
if not perfermance:
net_out = model(imgs) # [d1]
mask = net_out[0]
# ==============================================================================================================
# calculate distance map
imgs_ = imgs.cpu()
batchsize = imgs_.shape[0]
shape = list(imgs_.numpy().shape)
gts_onehot = make_one_hot(gts.unsqueeze(1).long(), num_classes=4)
shape[1] = 4
dist_onebatch = np.zeros(shape[1:])
dis_all = np.zeros(shape)
for i in range(batchsize):
for c in range(num_class):
dst, labels = cv2.distanceTransformWithLabels(
gts_onehot[i, c, ...].numpy().astype(np.uint8),
cv2.DIST_L2,
cv2.DIST_MASK_PRECISE,
labelType=cv2.DIST_LABEL_PIXEL)
# dst = list(dst)
dist_onebatch[c] = dst / np.max(dst) if np.max(
dst) else dst
dis_all[i] = dist_onebatch
dis_all_train = torch.from_numpy(dis_all).float().to(device)
# print(dis_all_train.shape)
# ==============================================================================================================
# ==============================================================================================================
# calculate mask gradient
mask_ = torch.softmax(mask, dim=1) # softmax
mask_ = mask_.cpu()
batchsize = mask_.detach().shape[0]
shape = list(mask_.detach().numpy().shape)
grad_onebatch = np.zeros(shape[1:])
grad_all = np.zeros(shape)
for i in range(batchsize):
for c in range(num_class):
cv2.imwrite('mask_i_c.png', mask_[i, c,
...].detach().numpy())
mask_i_c = cv2.imread('mask_i_c.png', 0)
os.remove('mask_i_c.png')
laplacian = cv2.Laplacian(mask_i_c, cv2.CV_64F)
# dst = list(dst)
grad_onebatch[c] = laplacian
grad_all[i] = grad_onebatch
grad_all_train = torch.from_numpy(grad_all).float().to(device)
# print(dis_all_train.shape)
# ==============================================================================================================
ce_loss = torch.nn.CrossEntropyLoss()(mask, gts.long())
smooth_loss = smooth_dist_loss(mask_true=make_one_hot(
gts.unsqueeze(1).long(), num_classes=4).to(device),
mask_pred=mask,
grad_pred=grad_all_train,
dist_true=dis_all_train)
loss = ce_loss + smooth_loss
tb_dict.update({"CE_loss": ce_loss.item()})
tb_dict.update({"smooth_loss": smooth_loss.item()})
tb_dict.update({"total_loss": loss.item()})
disp_dict.update({"CE_loss": ce_loss.item()})
disp_dict.update({"smooth_loss": smooth_loss.item()})
disp_dict.update({"total_loss": loss.item()})
if perfermance:
gts_ = gts.unsqueeze(1)
model.eval()
net_out_eval = model(imgs)[0]
net_out = F.softmax(net_out_eval, dim=1)
_, preds = torch.max(net_out, 1)
preds = preds.unsqueeze(1)
cal_perfer(make_one_hot(preds, num_class),
make_one_hot(gts_.long(), num_class), tb_dict)
model.train()
return ModelReturn(loss, tb_dict, disp_dict)
import cv2
import numpy as np
# 이미지를 읽어서 바이너리 스케일로 변환
img = cv2.imread('../img/2full_body.jpg')
imggray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
_, biimg = cv2.threshold(imggray, 127, 255, cv2.THRESH_BINARY_INV)
# 거리 변환 ---①
dst, labels = cv2.distanceTransformWithLabels(biimg, cv2.DIST_L2, 5)
print(np.unique(labels))
# 거리 값을 0 ~ 255 범위로 정규화 ---②
dst = (dst / (dst.max() - dst.min()) * 255).astype(np.uint8)
# 거리 값에 쓰레시홀드로 완전한 뼈대 찾기 ---③
skeleton = cv2.adaptiveThreshold(dst, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, 7, -3)
# 결과 출력
skeleton = cv2.cvtColor(skeleton, cv2.COLOR_GRAY2BGR)
for i in np.unique(labels):
skeleton[labels == i] = [int(j) for j in np.random.randint(0, 255, 3)]
cv2.imshow('origin', img)
cv2.imshow('dist', dst)
cv2.imshow('skel', skeleton)
cv2.waitKey(0)
cv2.destroyAllWindows()
def imageProcessing():
with graph.as_default():
with picamera.PiCamera() as camera:
with picamera.array.PiRGBArray(camera) as stream:
# カメラの解像度を320x240にセット
camera.resolution = (320, 240)
# カメラのフレームレートを15fpsにセット
camera.framerate = 15
# ホワイトバランスをfluorescent(蛍光灯)モードにセット
camera.awb_mode = 'fluorescent'
while True:
# stream.arrayにBGRの順で映像データを格納
camera.capture(stream, 'bgr', use_video_port=True)
# 映像データをHSV形式に変換
hsv = cv2.cvtColor(stream.array, cv2.COLOR_BGR2HSV)
# HSV形式からHチャンネルとSチャンネルの画像を得る
hsv_channels = cv2.split(hsv)
h_channel = hsv_channels[0]
s_channel = hsv_channels[1]
# Hチャンネルを平滑化
h_binary = cv2.GaussianBlur(h_channel, (5, 5), 0)
# Hチャンネルの二値化画像を作成
# hmin～hmaxの範囲を255（白）に、それ以外を0（黒）に
ret, h_binary = cv2.threshold(h_binary, hmax, 255,
cv2.THRESH_TOZERO_INV)
ret, h_binary = cv2.threshold(h_binary, hmin, 255,
cv2.THRESH_BINARY)
# smin～255の範囲を255（白）に、それ以外を0に（黒）に
ret, s_binary = cv2.threshold(s_channel, smin, 255,
cv2.THRESH_BINARY)
# HチャンネルとSチャンネルの二値化画像のANDをとる
# HチャンネルとSチャンネルの両方で255（白）の領域のみ白となる
hs_and = h_binary & s_binary
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
if CVversion == 2:
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - hs_and, cv2.cv.CV_DIST_L2, 5)
else:
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - hs_and, cv2.DIST_L2, 5)
img_label = np.uint8(img_label) & hs_and
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最大の白領域のみを残す
hand = np.uint8(img_label == m) * 255
# 最大の白領域からscikit-learnに入力するためのベクトルを取得
hand_vector = getImageVector(hand)
# 学習済のニューラルネットワークから結果を取得
X = np.array(hand_vector)
if K.image_data_format() == 'channels_first':
X = X.reshape(X.shape[0], 1, img_rows, img_cols)
input_shape = (1, img_rows, img_cols)
else:
X = X.reshape(X.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)
result = model.predict_classes(X, verbose=0)
# 分類結果をrecognizedHandに格納
global recognizedHand
recognizedHand = result[0]
# 手と判定されている領域を表示
cv2.imshow('hand', hand)
# waitを入れる
key = cv2.waitKey(1)
if appliStop == True:
break
# streamをリセット
stream.seek(0)
stream.truncate()
cv2.destroyAllWindows()
app.jankenStop()
app.quit()
def watershedSegmentation(wb, dataCube, plantMaterial, lighting, destination, datacubeName):
"""
Segements the plant material with watershed segmentation algorithm,
saves the segmented data in CSV files.
:param wb: (Spectronon workbench)
:param dataCube: (numpy array) BIL Datacube
:param plantMaterial: (numpy array) Locations of the plant material
:param lighting: (string) Illumination description string
:param destination: (string) Directory in which to save the data
:param datacubeName: (string) Name of the current data cube
:return: (None)
"""
lines, samples, bands = dataCube.shape
plantMaterialGray = plantMaterial
plantMaterial = cv2.merge((plantMaterialGray, plantMaterialGray, plantMaterialGray))
# Remove noise from the image
kernel = np.ones((3,3), np.uint8)
opening = cv2.morphologyEx(plantMaterialGray, cv2.MORPH_OPEN, kernel)
# Locate the for-sure background area
sure_bg = cv2.dilate(opening, kernel)
# Finding the for-sure foreground area
dist_transform, labels = cv2.distanceTransformWithLabels(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.4*dist_transform.max(), 255, 0)
sure_fg = cv2.morphologyEx(sure_fg, cv2.MORPH_CLOSE, kernel)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers+1
# Now, mark the region of unknown with zero. We will have the segmented "leaves"
# present in the image
markers[unknown==255] = 0
markers = cv2.watershed(plantMaterial, markers)
numMarkers = max(markers.flatten())
plt.imshow(markers)
plt.show()
saveData = wb.postQuestion("Continue with saving the segmented data?",
title="Segmentation Okay?")
if not saveData:
return
# Save the data for each leaf in a separate csv file (skip the background)
for leaf in xrange(2, numMarkers):
outputData = dataCube[markers==leaf, :]
if len(outputData.flatten()) / 240 < 2:
break
filename = datacubeName + '_' + lighting + '_' + str(leaf).zfill(4) + ".csv"
fileLocation = os.path.join(destination, filename)
np.savetxt(fileLocation, outputData, delimiter=',')
def model_fn(model,
data,
criterion,
perfermance=False,
vis=False,
device="cuda",
epoch=0,
num_class=4,
w=None):
# imgs, gts, _ = data
imgs, gts = data[:2]
imgs = imgs.to(device)
gts = torch.squeeze(gts, 1).to(device)
gts = gts.to(device)
loss = 0
tb_dict = {}
disp_dict = {}
if not perfermance:
net_out = model(imgs) # [d1]
mask = net_out[0]
dist = net_out[1]
# ==============================================================================================================
# calculate distance map
imgs_ = imgs.cpu()
batchsize = imgs_.shape[0]
shape = list(imgs_.numpy().shape)
gts_onehot = make_one_hot(gts.unsqueeze(1).long(), num_classes=4)
shape[1] = 4
dist_onebatch = np.zeros(shape[1:])
dis_all = np.zeros(shape)
for i in range(batchsize):
for c in range(num_class):
dst, labels = cv2.distanceTransformWithLabels(
gts_onehot[i, c, ...].numpy().astype(np.uint8),
cv2.DIST_L2,
cv2.DIST_MASK_PRECISE,
labelType=cv2.DIST_LABEL_PIXEL)
# dst = list(dst)
dist_onebatch[c] = dst
dis_all[i] = dist_onebatch
dis_all_train = torch.from_numpy(dis_all).float().to(device)
# print(dis_all_train.shape)
# ==============================================================================================================
ce_loss = torch.nn.CrossEntropyLoss()(mask, gts.long())
dist_loss = torch.nn.MSELoss()(
dist, dis_all_train) / dist.flatten().size()[0]
# dc_loss = DiceLoss()(net_out,make_one_hot(gts.unsqueeze(1).long(),num_classes=4).to(device))
loss = ce_loss + dist_loss
tb_dict.update({"CE_loss": ce_loss.item()})
tb_dict.update({"dist_loss": dist_loss.item()})
tb_dict.update({"total_loss": loss.item()})
disp_dict.update({"CE_loss": ce_loss.item()})
disp_dict.update({"dist_loss": dist_loss.item()})
disp_dict.update({"total_loss": loss.item()})
if perfermance:
gts_ = gts.unsqueeze(1)
model.eval()
net_out_eval = model(imgs)[0]
net_out = F.softmax(net_out_eval, dim=1)
_, preds = torch.max(net_out, 1)
preds = preds.unsqueeze(1)
cal_perfer(make_one_hot(preds, num_class),
make_one_hot(gts_.long(), num_class), tb_dict)
model.train()
return ModelReturn(loss, tb_dict, disp_dict)
def loadtd(self, idx):
# generate mask
image = cv2.imread('{}/MSRA-TD500/train/text_image/{}'.format(self.data_dir, idx), 1)
height = image.shape[0]
width = image.shape[1]
mask = np.zeros(image.shape, dtype=np.uint8)
bbox = ReadGt_td('{}/MSRA-TD500/train/gt/'.format(self.data_dir), idx[:-4])
for bboxi in range(len(bbox)):
pts, hard = GetPts_td(bbox, bboxi)
if hard == 1: continue
cv2.fillPoly(mask,[pts],(255,255,255),1)
# random crop parameters
attempt = 0
min_overlap = [0.1, 0.3, 0.5, 0.7]
while attempt < 10:
patch_h0, patch_w0, patch_h, patch_w = random_crop(mask)
overlap = mask[patch_h0:patch_h0+patch_h, patch_w0:patch_w0+patch_w] > 0
random_idx = np.random.randint(0, 4)
if np.sum(overlap) >= np.sum(mask>0)*min_overlap[random_idx]:
break
attempt += 1
if attempt == 10:
patch_h = min(height, width)
patch_w = patch_h
patch_h0 = (height-patch_h) // 2
patch_w0 = (width-patch_w) // 2
# random crop & resize
image = image[patch_h0:patch_h0+patch_h, patch_w0:patch_w0+patch_w, :]
image = cv2.resize(image, (self.patch_size, self.patch_size), interpolation=cv2.INTER_LINEAR)
# random rotate
prob = np.random.uniform(0,1)
if prob <= 0.2:
rtimes = 1
elif prob >= 0.8:
rtimes = 3
else:
rtimes = 0
for rcount in range(rtimes):
image = np.rot90(image)
# cv2.imwrite('train_input/{}'.format(idx),image)
# normalization
image = image.astype(np.float32)
image -= self.mean
image = image.transpose((2,0,1))
# compute vec
accumulation = np.zeros((3, self.patch_size, self.patch_size), dtype=np.float32)
for bboxi in range(len(bbox)):
pts, hard = GetPts_td(bbox, bboxi)
seg = np.zeros(mask.shape, dtype=np.uint8)
cv2.fillPoly(seg,[pts],(255,255,255),1)
segPatch = seg[patch_h0:patch_h0+patch_h, patch_w0:patch_w0+patch_w, :]
for rcount in range(rtimes):
segPatch = np.rot90(segPatch)
if segPatch.max()!=255: continue
segPatch = cv2.resize(segPatch, (self.patch_size, self.patch_size), interpolation=cv2.INTER_LINEAR)
cv2.rectangle(segPatch,(0,0),(self.patch_size-1,self.patch_size-1),(0,0,0),1)
img = cv2.cvtColor(segPatch,cv2.COLOR_BGR2GRAY)
img = (img > 128).astype(np.uint8)
dst, labels = cv2.distanceTransformWithLabels(img, cv2.cv.CV_DIST_L2, cv2.cv.CV_DIST_MASK_PRECISE, labelType=cv2.DIST_LABEL_PIXEL)
index = np.copy(labels)
index[img > 0] = 0
place = np.argwhere(index > 0)
nearCord = place[labels-1,:]
x = nearCord[:, :, 0]
y = nearCord[:, :, 1]
nearPixel = np.zeros((2, self.patch_size, self.patch_size))
nearPixel[0,:,:] = x
nearPixel[1,:,:] = y
grid = np.indices(img.shape)
grid = grid.astype(float)
diff = grid - nearPixel
dist = np.sqrt(np.sum(diff**2, axis = 0))
direction = np.zeros((3, self.patch_size, self.patch_size), dtype=np.float32)
direction[0,img > 0] = np.divide(diff[0,img > 0], dist[img > 0])
direction[1,img > 0] = np.divide(diff[1,img > 0], dist[img > 0])
if hard == 0:
direction[2,img > 0] = bboxi+1
accumulation[0,img > 0] = 0
accumulation[1,img > 0] = 0
accumulation[2,img > 0] = 0
accumulation = accumulation + direction
vec = np.stack((accumulation[0], accumulation[1]))
# compute weight
weight = np.zeros((self.patch_size, self.patch_size), dtype=np.float32)
posRegion = accumulation[2]>0
posCount = np.sum(posRegion)
if posCount != 0:
bboxRemain = 0
for bboxi in range(len(bbox)):
overlap_bboxi = accumulation[2]==(bboxi+1)
overlapCount_bboxi = np.sum(overlap_bboxi)
if overlapCount_bboxi == 0: continue
bboxRemain = bboxRemain+1
bboxAve = float(posCount)/bboxRemain
for bboxi in range(len(bbox)):
overlap_bboxi = accumulation[2]==(bboxi+1)
overlapCount_bboxi = np.sum(overlap_bboxi)
if overlapCount_bboxi == 0: continue
pixAve = bboxAve/overlapCount_bboxi
weight = weight*(~overlap_bboxi) + pixAve*overlap_bboxi
weight = weight[np.newaxis, ...]
return image, vec, weight
marker2.color.g = 0.0 marker2.color.b = 1.0 marker2.pose.orientation.w = 1.0 marker2.lifetime = rospy.Duration() ######################################### print "precomputing distance field .." t1 = time.time() #outsource to cython obstaclemap = util.get_obstaclemap(mapdata) (distance_field, labels) = cv2.distanceTransformWithLabels(obstaclemap, cv2.DIST_L2, cv2.DIST_MASK_PRECISE, labelType=cv2.DIST_LABEL_PIXEL) distance_field = np.array(distance_field, dtype=float) t2 = time.time() print "computed distance field in ", (t2-t1) pf = util.ParticleFilter(mapdata, distance_field, 15000) with open(logfile) as f: i = 0 for line in f:
def run(controller):
maps_initialized = False
count = 308
while (True):
#sleepで更新速度を制御
time.sleep(0.4)
frame = controller.frame()
image = frame.images[0]
if image.is_valid:
if not maps_initialized:
left_coordinates, left_coefficients = convert_distortion_maps(
frame.images[0])
right_coordinates, right_coefficients = convert_distortion_maps(
frame.images[1])
maps_initialized = True
undistorted_left = undistort(image, left_coordinates,
left_coefficients, 400, 400)
undistorted_right = undistort(image, right_coordinates,
right_coefficients, 400, 400)
#画像を２値化（白黒に処理）
ret, hand = cv2.threshold(undistorted_right, 70, 255,
cv2.THRESH_BINARY)
my_hand = hand[80:320, 40:360]
# 以下、最も広い白領域のみを残すための計算
# まず、白領域の塊（クラスター）にラベルを振る
img_dist, img_label = cv2.distanceTransformWithLabels(
255 - my_hand, cv2.DIST_L2, 5)
img_label = np.uint8(img_label) & my_hand
# ラベル0は黒領域なので除外
img_label_not_zero = img_label[img_label != 0]
# 最も多く現れたラベルが最も広い白領域のラベル
if len(img_label_not_zero) != 0:
m = stats.mode(img_label_not_zero)[0]
else:
m = 0
# 最も広い白領域のみを残す
this_hand = np.uint8(img_label == m) * 255
#膨張
#kernel = np.ones((5,5),np.uint8)
#this_hand = cv2.dilate(this_hand,kernel,iterations = 1)
#hand_color = cv2.cvtColor(this_hand,cv2.COLOR_GRAY2BGR)
# 輪郭を抽出
contours, hierarchy = cv2.findContours(this_hand,
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
cnt = contours[0]
#重心を求める
mu = cv2.moments(this_hand, False)
mx, my = int(mu["m10"] / mu["m00"]), int(mu["m01"] / mu["m00"])
#手首の位置を求める
frame = controller.frame()
righthand = frame.hands.rightmost
arm = righthand.arm
i_box = frame.interaction_box
normalized_tip = i_box.normalize_point(arm.wrist_position)
app_x = 160 * normalized_tip.x + 80
app_y = 120 * (normalized_tip.z) + 60
app = (int(app_x), int(app_y))
#重心と手首の位置から回転させる
angle = 90 + math.degrees(math.atan2(my - app[1], mx - app[0]))
trans = cv2.getRotationMatrix2D((mx, my), angle, 1.0)
this_hand = cv2.warpAffine(this_hand, trans, (360, 240))
# 得られた二値化画像を画面に表示
cv2.imshow('hand', this_hand)
#cv2.imshow('hand', undistorted_right)
cv2.imwrite("HandImages2/img_two{0:03d}.png".format(count),
this_hand)
count += 1
if cv2.waitKey(1) & 0xFF == ord('q'):
break
