def compare_with(self, image2, downsample_ratio = 35.0):
ret = {'scale': float(downsample_ratio), 'offsets': [0, 0]}
digipal_images = [self, Image.from_digipal_image(image2)]
# Get PIL images of the thumbnails
ims = []
for image in digipal_images:
ims.append(image.get_pil_img(1.0/downsample_ratio, cache=True, grey=True))
# Make sure the largest image is ims[0]
sgn = -1
reverse_order = False
if ims[0].size[0] < ims[1].size[0]:
sgn = 1
ims = ims[::-1]
digipal_images = digipal_images[::-1]
reverse_order = True
# -------------------
# Step 1: Find the approximate crop offset using thumbnails
# -------------------
offsets = self.find_offsets_from_pil_images(ims[0], ims[1])
if offsets is None:
return ret
# convert the offset from the thumbnail size to the full size
ret['offsets'] = [offset * downsample_ratio * sgn for offset in offsets]
return ret
def replace_image_and_update_annotations(self, offsets, new_image):
ret = True
old_image_size = self.get_img_size()
new_image = Image.from_digipal_image(new_image)
new_image_size = new_image.get_img_size()
self.iipimage = new_image.iipimage
self.save()
annotations = self.annotation_set.filter().distinct()
if annotations.count():
for annotation in annotations:
annotation.geo_json = self.get_uncropped_geo_json(annotation, offsets, new_image_size, old_image_size)
annotation.save()
return ret
def replace_image_and_update_annotations(self, offsets, new_image):
ret = True
old_image_size = self.get_img_size()
new_image = Image.from_digipal_image(new_image)
new_image_size = new_image.get_img_size()
self.iipimage = new_image.iipimage
self.save()
annotations = self.annotation_set.filter().distinct()
if annotations.count():
for annotation in annotations:
print 'translate annotation #%s' % annotation.id
annotation.geo_json = self.get_uncropped_geo_json(
annotation, offsets, new_image_size, old_image_size)
annotation.save()
return ret
def compare_with(self, image2, downsample_ratio=35.0):
ret = {'scale': float(downsample_ratio), 'offsets': [0, 0]}
digipal_images = [self, Image.from_digipal_image(image2)]
# Get PIL images of the thumbnails
ims = []
for image in digipal_images:
ims.append(
image.get_pil_img(1.0 / downsample_ratio,
cache=True,
grey=True))
# Make sure the largest image is ims[0]
sgn = -1
reverse_order = False
if ims[0].size[0] < ims[1].size[0]:
sgn = 1
ims = ims[::-1]
digipal_images = digipal_images[::-1]
reverse_order = True
# -------------------
# Step 1: Find the approximate crop offset using thumbnails
# -------------------
offsets = self.find_offsets_from_pil_images(ims[0], ims[1])
if offsets is None:
return ret
# convert the offset from the thumbnail size to the full size
ret['offsets'] = [
offset * downsample_ratio * sgn for offset in offsets
]
return ret
def find_image_offset(self, image2, downsample_ratio = 35.0):
'''
Find and returns the crop offset and two sample annotations.
return = {'offsets': [x,y], 'annotations': [a1, a2]}
(x, y), how much do we have to move image1 to match image2
a1 is the URL of one annotation, a2 the URL of the matching
annotation in the other image
The offset is found using this tow-steps method:
1. obtains a thumbnail of the images and use PIL to
scan all possible offsets and return the best match.
Note that this is an approximate offset due to
the downsampling made for creating the thumbnails.
2. the approximate is then refined by applying the same
PIL matching operation on one annotation cutout from
one image and the region where we expect to find it
in the other image.
The smaller the thumbnail in step 1, the larger the
search region in this step.
downsample_ratio specifies how small the thumbnail is compared
to the original. Value above 35 may cause problems with our
images as the approximation in the first stage will be excessive.
'''
ret = {'offsets': [0,0], 'annotations': []}
if self.annotation_set.count() + image2.annotation_set.count() == 0:
return ret
digipal_images = [self, Image.from_digipal_image(image2)]
# Get PIL images of the thumbnails
ims = []
for image in digipal_images:
ims.append(image.get_pil_img(1.0/downsample_ratio, cache=True, grey=True))
# Make sure the largest image is ims[0]
reverse_order = False
if ims[0].size[0] < ims[1].size[0]:
ims = ims[::-1]
digipal_images = digipal_images[::-1]
reverse_order = True
# -------------------
# Step 1: Find the approximate crop offset using thumbnails
# -------------------
best = self.find_offsets_from_pil_images(ims[0], ims[1])
if best is None:
return ret
# -------------------
# Step 2: Compare an annotation from one image with the containing region in the other image
# -------------------
# convert the offset from the thumbnail size to the full size
ret['offsets'] = [best[0] * downsample_ratio, best[1] * downsample_ratio]
# get one annotation (ann)
for annotation_image_index in range(0, 2):
anns = digipal_images[annotation_image_index].annotation_set.all().order_by('?')
#anns = digipal_images[annotation_image_index].annotation_set.all().order_by('id')
if anns.count():
ann = anns[0]
break
# Calculate the search region (region_search):
# start from the coordinates of the annotation in the first image (region)
region = self.get_annotation_coordinates(ann, digipal_images[annotation_image_index])
region_search = []
sign = 1
if annotation_image_index == 0:
sign = -1
for pair in region:
region_search.append([pair[0] + ret['offsets'][0] * sign, pair[1] + ret['offsets'][1] * sign])
# extend the search region by downsample_ratio/2 pixels in every direction
# (due to approximation of the first step - every pixel in thumbnail =
# downsample_ratio pixels in the original)
#
# Now add a bit more to that because of the rounding errors when asking for the
# thumbnail dimensions and other possible source of perturbation.
# We take 50% (* 1.5) safety margin more on each side
#
safety_margin = 2.2
search_margin = int((downsample_ratio / 2.0) * safety_margin + 0.5)
region_search = [
[region_search[0][0] - search_margin, region_search[0][1] - search_margin],
[region_search[1][0] + search_margin, region_search[1][1] + search_margin],
]
# iip img server only accepts relative coordinates, so convert them into relatives
region_search_relative = digipal_images[1-annotation_image_index].get_relative_coordinates(region_search)
# Get the two images (annotation and search region) as PIL objects
ann_im = self.get_pil_img_from_url(ann.get_cutout_url(False, True), grey=True, cache=True)
rgn_im = digipal_images[1-annotation_image_index].get_pil_img(
query='&RGN=%1.6f,%1.6f,%1.6f,%1.6f&QLT=100&CVT=JPG' % (region_search_relative[0][0], region_search_relative[0][1], region_search_relative[1][0], region_search_relative[1][1])
, grey=True
, cache=True
)
# Find the offsets by systematically looking for a match of the annotation within
# the search region
offsets = self.find_offsets_from_pil_images(rgn_im, ann_im)
# Now add local offset to the global estimate
ret['offsets'] = [offsets[0] + ret['offsets'][0] - search_margin, offsets[1] + ret['offsets'][1] - search_margin]
# Get the URL of the reference annotation and the matching one in the other image
ret_relative = digipal_images[1-annotation_image_index].get_relative_coordinates([[region[0][0] + ret['offsets'][0], region[0][1] + ret['offsets'][1]], [region[1][0] + ret['offsets'][0], region[1][1] + ret['offsets'][1]]])
ann_img_2_url = digipal_images[1-annotation_image_index].iipimage.full_base_url + '&RGN=%1.6f,%1.6f,%1.6f,%1.6f&QLT=100&CVT=JPG' % (ret_relative[0][0], ret_relative[0][1], ret_relative[1][0], ret_relative[1][1])
ret['annotations'] = [ann.get_cutout_url(False, True), ann_img_2_url]
# Adjust the crop signs
if not reverse_order:
ret['offsets'] = [-ret['offsets'][0], -ret['offsets'][1]]
return ret
