def main(quantMat, path):
#quantMat = np.full([8,8], 1)
#Read Image
orgImage = imt.ImageTransform()
orgRGB = orgImage.readImage(path)
#Reduce resolution
imgPIL = orgImage.rgb2img(orgRGB)
halfImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 2)
quartImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 4)
'''Hierachical Encoding'''
#Encode quartImg
F_info_quart, orgShape_quart = coding.encode(np.array(quartImg), quantMat)
#print orgShape_quart
#Decode QuartImg
recRGB_quart, recImg2show_quart = coding.decode(F_info_quart, quantMat,
orgShape_quart)
#recImg2show_quart.show()
#Encode half diff Image
hierVar = hr.Hier(quantMat)
halfdiffF_info, halfdifforgShape, hafldiffrecRGB, halfdiffrecImg2show = hierVar.hierEncode(
np.array(quartImg), np.array(halfImg))
#Decode half
recRGB_halfdiff, recImg2show_halfdiff = coding.decode(
halfdiffF_info, quantMat, halfdifforgShape)
fRecoverHalf, fRecoverHalfImg = hierVar.hireDecode(np.array(quartImg),
recRGB_halfdiff)
#fRecoverHalfImg.show()
#fRecover, fRecoverImg = hierVar.hireDecode(fRecoverHalf, diffF_info, difforgShape)
#Encode whole diff Image
#imgPIL.show()
diffF_info, difforgShape, diffrecRGB, diffrecImg2show = hierVar.hierEncode(
np.array(fRecoverHalfImg), np.array(imgPIL))
#Decode whole image
recRGB_diff, recImg2show_diff = coding.decode(diffF_info, quantMat,
difforgShape)
fRecover, fRecoverImg = hierVar.hireDecode(np.array(fRecoverHalfImg),
recRGB_diff)
#fRecoverHalfImg.show()
#fRecoverImg.show()
return [
imgPIL, fRecoverImg, halfdiffrecImg2show, diffrecImg2show,
fRecoverHalfImg, recImg2show_quart
]
def detect_markers(img):
width, height, _ = img.shape
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
edges = cv2.Canny(gray, 10, 100)
contours, hierarchy = cv2.findContours(edges.copy(), cv2.RETR_TREE,
cv2.CHAIN_APPROX_NONE)[-2:]
# We only keep the long enough contours
min_contour_length = min(width, height) / 50
contours = [
contour for contour in contours if len(contour) > min_contour_length
]
warped_size = 49
canonical_marker_coords = array(
((0, 0), (warped_size - 1, 0), (warped_size - 1, warped_size - 1),
(0, warped_size - 1)),
dtype='float32')
markers_list = []
transform_list = []
for contour in contours:
approx_curve = cv2.approxPolyDP(contour, len(contour) * 0.01, True)
if not (len(approx_curve) == 4 and cv2.isContourConvex(approx_curve)):
continue
sorted_curve = array(cv2.convexHull(approx_curve, clockwise=False),
dtype='float32')
persp_transf = cv2.getPerspectiveTransform(sorted_curve,
canonical_marker_coords)
transform_list.append(persp_transf)
warped_img = cv2.warpPerspective(img, persp_transf,
(warped_size, warped_size))
warped_gray = cv2.cvtColor(warped_img, cv2.COLOR_BGR2GRAY)
_, warped_bin = cv2.threshold(warped_gray, 127, 255, cv2.THRESH_BINARY)
marker = warped_bin.reshape([
MARKER_SIZE, warped_size / MARKER_SIZE, MARKER_SIZE,
warped_size / MARKER_SIZE
])
marker = marker.mean(axis=3).mean(axis=1)
marker[marker < 127] = 0
marker[marker >= 127] = 1
try:
marker = validate_and_turn(marker)
hamming_code = extract_hamming_code(marker)
marker_id = int(decode(hamming_code), 2)
markers_list.append(
HammingMarker(id=marker_id, contours=approx_curve))
except ValueError:
continue
return markers_list
def main(quantMat, path): #quantMat = np.full([8,8], 1) #Read Image orgImage = imt.ImageTransform() orgRGB = orgImage.readImage(path) #Reduce resolution imgPIL = orgImage.rgb2img(orgRGB) halfImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 2) quartImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 4) '''Hierachical Encoding''' #Encode quartImg F_info_quart, orgShape_quart = coding.encode(np.array(quartImg), quantMat) #print orgShape_quart #Decode QuartImg recRGB_quart, recImg2show_quart = coding.decode(F_info_quart, quantMat, orgShape_quart) #recImg2show_quart.show() #Encode half diff Image hierVar = hr.Hier(quantMat) halfdiffF_info, halfdifforgShape, hafldiffrecRGB, halfdiffrecImg2show = hierVar.hierEncode(np.array(quartImg), np.array(halfImg)) #Decode half recRGB_halfdiff, recImg2show_halfdiff = coding.decode(halfdiffF_info, quantMat, halfdifforgShape) fRecoverHalf, fRecoverHalfImg = hierVar.hireDecode(np.array(quartImg), recRGB_halfdiff) #fRecoverHalfImg.show() #fRecover, fRecoverImg = hierVar.hireDecode(fRecoverHalf, diffF_info, difforgShape) #Encode whole diff Image #imgPIL.show() diffF_info, difforgShape, diffrecRGB, diffrecImg2show = hierVar.hierEncode(np.array(fRecoverHalfImg), np.array(imgPIL)) #Decode whole image recRGB_diff, recImg2show_diff = coding.decode(diffF_info, quantMat, difforgShape) fRecover, fRecoverImg = hierVar.hireDecode(np.array(fRecoverHalfImg), recRGB_diff) #fRecoverHalfImg.show() #fRecoverImg.show() return [imgPIL, fRecoverImg, halfdiffrecImg2show ,diffrecImg2show, fRecoverHalfImg ,recImg2show_quart]
def hierEncode(self, f_half, f): highResShape = np.shape(f) E_f_half = imt.ImageTransform.imresize(self.IMT.rgb2img(f_half), highResShape, 1) E_f_half_RGB = np.array(E_f_half) diffRGB = f.astype(int) - E_f_half_RGB.astype(int) diffRGB[diffRGB<0] = 1 diffRGB[diffRGB>255] = 255 diffRGB = diffRGB.astype(np.uint8) F_info, orgShape = coding.encode(np.array(diffRGB), self.quantMat) recRGB, recImg2show = coding.decode(F_info, self.quantMat, orgShape) return F_info, orgShape, recRGB, recImg2show
def hierEncode(self, f_half, f):
highResShape = np.shape(f)
E_f_half = imt.ImageTransform.imresize(self.IMT.rgb2img(f_half),
highResShape, 1)
E_f_half_RGB = np.array(E_f_half)
diffRGB = f.astype(int) - E_f_half_RGB.astype(int)
diffRGB[diffRGB < 0] = 1
diffRGB[diffRGB > 255] = 255
diffRGB = diffRGB.astype(np.uint8)
F_info, orgShape = coding.encode(np.array(diffRGB), self.quantMat)
recRGB, recImg2show = coding.decode(F_info, self.quantMat, orgShape)
return F_info, orgShape, recRGB, recImg2show
def detect_markers(img):
width, height, _ = img.shape
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
edges = cv2.Canny(gray, 10, 100)
contours, hierarchy = cv2.findContours(edges.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)
# We only keep the long enough contours
min_contour_length = min(width, height) / 50
contours = [contour for contour in contours if len(contour) > min_contour_length]
warped_size = 49
canonical_marker_coords = array(((0, 0),
(warped_size - 1, 0),
(warped_size - 1, warped_size - 1),
(0, warped_size - 1)),
dtype='float32')
markers_list = []
for contour in contours:
approx_curve = cv2.approxPolyDP(contour, len(contour) * 0.01, True)
if not (len(approx_curve) == 4 and cv2.isContourConvex(approx_curve)):
continue
sorted_curve = array(cv2.convexHull(approx_curve, clockwise=False),
dtype='float32')
persp_transf = cv2.getPerspectiveTransform(sorted_curve, canonical_marker_coords)
warped_img = cv2.warpPerspective(img, persp_transf, (warped_size, warped_size))
warped_gray = cv2.cvtColor(warped_img, cv2.COLOR_BGR2GRAY)
_, warped_bin = cv2.threshold(warped_gray, 127, 255, cv2.THRESH_BINARY)
marker = warped_bin.reshape(
[MARKER_SIZE, warped_size / MARKER_SIZE, MARKER_SIZE, warped_size / MARKER_SIZE]
)
marker = marker.mean(axis=3).mean(axis=1)
marker[marker < 127] = 0
marker[marker >= 127] = 1
try:
marker = validate_and_turn(marker)
hamming_code = extract_hamming_code(marker)
marker_id = int(decode(hamming_code), 2)
markers_list.append(HammingMarker(id=marker_id, contours=approx_curve))
except ValueError:
continue
return markers_list
def evaluate(net, loader, thres=0.5, max_aabbs=None):
batch_imgs = []
batch_aabbs = []
loss = 0
for i in range(len(loader)):
# get batch
loader_item = loader[i]
with torch.no_grad():
y = net(loader_item.batch_imgs, apply_softmax=True)
y_np = y.to('cpu').numpy()
if loader_item.batch_gt_maps is not None:
loss += compute_loss(
y, loader_item.batch_gt_maps).to('cpu').numpy()
scale_up = 1 / compute_scale_down(WordDetectorNet.input_size,
WordDetectorNet.output_size)
metrics = BinaryClassificationMetrics(0, 0, 0)
for i in range(len(y_np)):
img_np = loader_item.batch_imgs[i, 0].to('cpu').numpy()
pred_map = y_np[i]
aabbs = decode(pred_map,
comp_fg=fg_by_cc(thres, max_aabbs),
f=scale_up)
h, w = img_np.shape
aabbs = [aabb.clip(AABB(0, w - 1, 0, h - 1))
for aabb in aabbs] # bounding box must be inside img
clustered_aabbs = cluster_aabbs(aabbs)
if loader_item.batch_aabbs is not None:
curr_metrics = binary_classification_metrics(
loader_item.batch_aabbs[i], clustered_aabbs)
metrics = metrics.accumulate(curr_metrics)
batch_imgs.append(img_np)
batch_aabbs.append(clustered_aabbs)
return EvaluateRes(batch_imgs, batch_aabbs, loss / len(loader), metrics)
class DatasetIAMSplit:
"""wrapper which provides a dataset interface for a split of the original dataset"""
def __init__(self, dataset, start_idx, end_idx):
assert start_idx >= 0 and end_idx <= len(dataset)
self.dataset = dataset
self.start_idx = start_idx
self.end_idx = end_idx
def __getitem__(self, idx):
return self.dataset[self.start_idx + idx]
def __len__(self):
return self.end_idx - self.start_idx
if __name__ == '__main__':
from visualization import visualize
from coding import encode, decode
import matplotlib.pyplot as plt
dataset = DatasetIAM(Path('../data'), (350, 350), (350, 350), caching=False)
img, gt = dataset[0]
gt_map = encode(img.shape, gt)
gt = decode(gt_map)
plt.imshow(visualize(img / 255 - 0.5, gt))
plt.show()
def check(s):
c = encode(s)
s2 = decode(c)
assert '\n' not in c[:-1]
assert c[-1] == '\n'
assert s == s2
def on_decode_button_clicked(self):
method = self.method_chooser.get()
code = self.code.get('1.0', 'end')[:-1]
text = coding.decode(method, code)
self.text.delete('1.0', 'end')
self.text.insert('1.0', text)
def load_line(line):
return pickle.loads(decode(line))
