def generate_positive_regions(image_dir, bbinfo_dir, modifiers_config=None, window_dims=None, min_size=(48,48)):
print 'generate_positive_regions:'
all_images = utils.list_images_in_directory(image_dir)
# Create the modifier generator:
# Note: Simply generate null modifiers if no modifier config is passed.
modifier_generator = itertools.repeat(utils.RegionModifiers())
if modifiers_config:
modifier_generator = utils.RegionModifiers.random_generator_from_config_dict(modifiers_config)
# Filter out images without bounding boxes:
bbinfo_map = utils.load_opencv_bounding_box_info_directory(bbinfo_dir, suffix='bbinfo')
source_images = [img_path for img_path in all_images if not utils.info_entry_for_image(bbinfo_map, img_path) is None]
# Extract image regions:
source_regions = []
for img_path in source_images:
rects = utils.info_entry_for_image(bbinfo_map, img_path)
for rect in rects:
# Reject small samples:
if rect.w < float(min_size[0]) or rect.h < float(min_size[1]):
continue
region = utils.ImageRegion(rect, img_path)
source_regions.append(region)
print 'Found {} source regions.'.format(len(source_regions))
# Generate an infinite list of samples:
while True:
# Randomise the order of source images:
# Note: Don't randomise if no modifiers are used.
if modifiers_config:
random.shuffle(source_regions)
# For each source region:
for reg in source_regions:
mod = modifier_generator.next()
# Enlarge to correct aspect ratio:
new_rect = reg.rect
if window_dims:
aspect = window_dims[0] / float(window_dims[1])
new_rect = new_rect.enlarge_to_aspect(aspect)
imsize = utils.get_image_dimensions(reg.fname)
if not new_rect.lies_within_frame(imsize):
# print 'bad', new_rect
continue
# else:
# print new_rect
# Assign a random modifier, and yield the region:
new_reg = utils.ImageRegion(new_rect, reg.fname, mod)
yield new_reg
# If we're not using modifiers, only use each region once:
if not modifiers_config:
break
def generate_negative_regions_with_exclusions(bak_img_dir, exl_info_map, window_dims, modifiers_config=None):
print 'generate_negative_regions_with_exclusions:'
all_images = utils.list_images_in_directory(bak_img_dir)
if len(all_images) == 0:
raise ValueError('The given directory \'{}\' contains no images.'.format(bak_img_dir))
print ' Found {} images.'.format(len(all_images))
all_images = [img_path for img_path in all_images if not utils.info_entry_for_image(exl_info_map, img_path) is None]
print ' Found {} images with exclusion info.'.format(len(all_images))
if len(all_images) == 0:
raise ValueError('The given directory \'{}\' contains no images with exclusion info.'.format(bak_img_dir))
image_list = all_images
random.shuffle(image_list)
modifier_generator = itertools.repeat(utils.RegionModifiers())
if modifiers_config:
modifier_generator = utils.RegionModifiers.random_generator_from_config_dict(modifiers_config)
while True:
for img_path in image_list:
imsize = utils.get_image_dimensions(img_path)
for reg in generate_negative_regions_in_image_with_exclusions(img_path, exl_info_map, window_dims):
mod = modifier_generator.next()
# Assign a random modifier, and yield the region:
new_reg = utils.ImageRegion(reg.rect, reg.fname, mod)
yield new_reg
# Shuffle the image list:
random.shuffle(image_list)
def checkBoundingBoxes(img_paths, bbinfo_dir):
global_info = utils.load_opencv_bounding_box_info_directory(
bbinfo_dir, suffix='bbinfo')
# global_info = loadGlobalInfo(bbinfo_dir)
error_occurred = False
bad_images = []
warn_images = []
for img_path in img_paths:
imsize = None
try:
im = Image.open(img_path)
imsize = im.size
im.close()
except Exception as e:
print "Error opening image {}: {}".format(img_path, e)
error_occurred = True
continue
# Get bounding box:
rects = utils.info_entry_for_image(global_info, img_path)
for rect in rects:
if not rect.touches_frame_edge(imsize):
warn_images.append((img_path, rect, imsize))
if not rect.lies_within_frame(imsize):
bad_images.append((img_path, rect, imsize))
if warn_images:
print """
WARNING: Some bounding boxes touch the image edges.
An older version of this script defined bounding boxes differently
in which case the case where rect.x + rect.w == img_width would
indicate a bounding box extending outside of the image dimensions.
The newer KITTI scripts define bounding box size as the true pixel
size, making this test indicate a box touching the image edge, which
is fine.
If the OpenCV scripts fail, this may be the reason.
(Uncomment the following line to see the suspicious boxes)
"""
# pprint(warn_images)
if error_occurred or bad_images:
print 'EXITING DUE TO INVALID BOUNDING BOXES:'
pprint(bad_images)
sys.exit(1)
assert (False)
def checkBoundingBoxes(img_paths, bbinfo_dir):
global_info = utils.load_opencv_bounding_box_info_directory(bbinfo_dir, suffix='bbinfo')
# global_info = loadGlobalInfo(bbinfo_dir)
error_occurred = False
bad_images = []
warn_images = []
for img_path in img_paths:
imsize = None
try:
im = Image.open(img_path)
imsize = im.size
im.close()
except Exception as e:
print "Error opening image {}: {}".format(img_path, e)
error_occurred = True
continue
# Get bounding box:
rects = utils.info_entry_for_image(global_info, img_path)
for rect in rects:
if not rect.touches_frame_edge(imsize):
warn_images.append((img_path, rect, imsize))
if not rect.lies_within_frame(imsize):
bad_images.append((img_path, rect, imsize))
if warn_images:
print """
WARNING: Some bounding boxes touch the image edges.
An older version of this script defined bounding boxes differently
in which case the case where rect.x + rect.w == img_width would
indicate a bounding box extending outside of the image dimensions.
The newer KITTI scripts define bounding box size as the true pixel
size, making this test indicate a box touching the image edge, which
is fine.
If the OpenCV scripts fail, this may be the reason.
(Uncomment the following line to see the suspicious boxes)
"""
# pprint(warn_images)
if error_occurred or bad_images:
print 'EXITING DUE TO INVALID BOUNDING BOXES:'
pprint(bad_images)
sys.exit(1)
assert(False)
def write_img(img_path, dat_file, write_bbox=True):
rel_img_path = None
if os.path.isabs(img_path):
rel_img_path = os.path.relpath(img_path, abs_output_dir)
else:
rel_img_path = os.path.relpath(img_path, output_dir)
dat_line = rel_img_path
if write_bbox:
rects = utils.info_entry_for_image(global_info, img_path)
details = ' '.join([' '.join(map(str, rect.opencv_bbox)) for rect in rects])
dat_line = "{} {} {}".format(rel_img_path, len(rects), details)
# print 'dat_line:', dat_line
dat_file.write(dat_line)
def write_img(img_path, dat_file, write_bbox=True):
rel_img_path = None
if os.path.isabs(img_path):
rel_img_path = os.path.relpath(img_path, abs_output_dir)
else:
rel_img_path = os.path.relpath(img_path, output_dir)
dat_line = rel_img_path
if write_bbox:
rects = utils.info_entry_for_image(global_info, img_path)
details = ' '.join(
[' '.join(map(str, rect.opencv_bbox)) for rect in rects])
dat_line = "{} {} {}".format(rel_img_path, len(rects), details)
# print 'dat_line:', dat_line
dat_file.write(dat_line)
def generate_negative_regions_in_image_with_exclusions(img_path, exl_info_map, window_dims):
excl_info = utils.info_entry_for_image(exl_info_map, img_path)
# Only consider images with exclusions:
if excl_info is None:
return
imsize = utils.get_image_dimensions(img_path)
img_w, img_h = imsize
aspect = window_dims[0] / float(window_dims[1])
scale_step = 1.25
window_step = 0.5
max_scale = 0.5
min_w = 64
min_scale = min_w / float(img_w)
scale = min_scale
# Used to probabilistically reject a fraction of samples at each scale
# level. This decreases the difference in the number of samples selected
# from each scale level (otherwise, small scale levels would contain an
# overwhelmingly large number of sample windows).
def accept_sample(curr_scale):
f = (curr_scale - min_scale) / (max_scale - min_scale)
min_prob = 0.3**0.5
max_prob = 1.0
prob = min_prob*(1-f) + max_prob*(f)
return np.random.uniform() < prob*prob
# Probability of rejecting samples that are not close to exclusion regions.
# (Most images have large portions of sky and ground, and this prevents
# oversampling those areas)
far_reject_prob = 0.97
num_found = 0
while True:
w = int(round(img_w * scale))
h = int(round(img_w * scale / aspect))
sw = int(window_step*w)
sh = int(window_step*h)
# print 'scale, (w, h), (sw, sh):', scale, (w, h), (sw, sh)
if scale > max_scale:
return
scale *= scale_step
count = 0
for x in xrange(0, img_w, sw):
for y in xrange(0, img_h, sh):
if x + w > img_w or y + h > img_h:
break
# Randomly reject samples based on scale:
if not accept_sample(scale):
continue
rect = gm.PixelRectangle.from_opencv_bbox([x, y, w, h])
# Nudge rectangles to avoid exclusion regions:
intersecting = filter(rect.intersects_pixelrectangle, excl_info)
if len(intersecting) == 1:
excl = intersecting[0]
rect, offset = rect.moved_to_clear(excl, return_offset=True)
if abs(offset[0]) >= sw or abs(offset[1]) >= sh:
continue
rect = rect.translated((0,0), imsize)
# Ensure the rectangle does not intersect an exclusion region:
if not any((rect.intersects_pixelrectangle(pr) for pr in excl_info)):
# Determing whether the rectangle is close to an exclusion
# region:
is_close_to_exclusion = any((rect.distance_pixelrectangle(pr) < w for pr in excl_info))
# Prefer rectangles that are close to exclusion regions:
if not is_close_to_exclusion:
if np.random.uniform() < far_reject_prob:
continue
reg = utils.ImageRegion(rect, img_path)
num_found += 1
# import sys
# sys.stdout.write(str(num_found) + ',')
# sys.stdout.flush()
yield reg
count += 1
