def save_tiles(self, path, rows=None, cols=None, overlap=16):
if rows is None:
rows = self.nrows
if cols is None:
cols = rows
tiles = split_tiles(channels_first(self.data), (rows, cols), overlap)
for i, tile in enumerate(tiles):
fn = '{}_tile{:04}.npy'.format(path, i+1)
np.save(fn, tile)
def rectify_building(image, meta=None):
labels = pyfacades.models.driving_12x360x480.segment.process_strip(
channels_first(image))
non_rectangular = building = np.argmax(
labels, 0) == pyfacades.models.driving_12x360x480.model.BUILDING
h = pyfacades.rectify.Homography(image, mask=non_rectangular)
if meta is not None:
meta['labels'] = labels
meta['building'] = building
meta['homography'] = h
return h.rectified
def load_image(self, path):
self.data = imread(path)
self.path = path
self._load_image_mask()
self._mask_out_common_obstructions()
self._rectify_image()
self.driving_layers = driving.process_strip(
channels_first(self.rectified * 255))
self.facade_layers = i12.process_strip(
channels_first(self.rectified * 255))
self._create_sky_mask()
self._segment_windows()
self._segment_facade_edges()
facade_cuts = self._split_at_facade_edges()
facade_mask = self._create_facade_mask()
wall_colors = self._mask_out_wall_colors(facade_mask)
self.wall_colors = wall_colors
self.facade_candidates = self._find_facade_candidates(
wall_colors, facade_cuts)
def _mask_out_common_obstructions(self):
"""Mask out the sky and some common objects that can obstruct a facade
This is intended to be run prior to rectifications, since these things can prevent us from
correctly identifying the rectilinear features of a facade
"""
features = driving.process_strip(
channels_first(img_as_ubyte(self.data)))
if self.data_mask is not None:
occlusions = ~self.data_mask
else:
occlusions = np.zeros(self.data.shape[:2], dtype=np.bool)
occlusions |= driving.occlusion(features)
self.data_mask = ~occlusions
def eval():
p = ArgumentParser(default_config_files=[join(dirname(__file__), '.eval.cfg')])
p.add('--config', '-c', is_config_file=True, help="config file path")
p.add('--weights', type=str, help="path to the classifier weights")
p.add('--data', type=str, help="test data, a file with each line as an .npy filename")
p.add('--output', '-o', type=str, help="folder to hold outputs", default='.')
p.add('--path', type=str, action='append', help='search paths for files')
args = p.parse_args()
try:
os.makedirs(args.output)
except OSError:
pass
# If weights is None, this will default to the weights in the models folder,
# or the ones indicated by the environment variable
net = i12.net(args.weights)
files = [complete(p, args.path) for p in open(args.data).readlines()]
summary = OrderedDict()
for label in i12.LABELS:
summary[label] = OrderedDict()
summary[label].update(TP=0)
summary[label].update(TN=0)
summary[label].update(FP=0)
summary[label].update(FN=0)
summary[label].update(targets=0)
summary[label].update(candidates=0)
summary[label].update(hits=0)
for file_index, file in enumerate(files):
print(file_index+1, "of", len(files), ":", file)
cached_file = os.path.join(args.output, "metrics-{:04}.yml".format(file_index))
data = np.load(file)
rgb = channels_last(data[:3].astype(np.uint8))
if not os.path.isfile(os.path.join(args.output, 'file-{}.jpg'.format(file_index))):
skimage.io.imsave(os.path.join(args.output, 'file-{}.jpg'.format(file_index)), rgb)
all_expected = data[3:]
if not os.path.isfile(os.path.join(args.output, 'file-{}-windows.jpg'.format(file_index))):
skimage.io.imsave(os.path.join(args.output, 'file-{}-windows.jpg'.format(file_index)),
skimage.color.gray2rgb((all_expected[
pyfacades.models.independant_12_layers.model.WINDOW] == 2).astype(float)))
if not os.path.isfile(cached_file):
print("Calculating metrics for file", file)
#data = np.load(file)
#rgb = channels_last(data[:3].astype(np.uint8))
#all_expected = data[3:]
all_predicted = i12.process_strip(channels_first(rgb))
results = OrderedDict()
for label_index, label in enumerate(i12.LABELS):
if label == 'background': continue
# Skip the top row and the bottom row -- I gave those dummy labels to work around a caffe error
expected = all_expected[label_index][1:-1]
# Uggh! Since the 'UNKNOWN' label is removed, argmax gives the wrong values; I need 0->0, 1->2, 2->3
predicted = all_predicted.features[label_index].argmax(0)[1:-1]+1
predicted[predicted==LABEL_UNKNOWN] = LABEL_NEGATIVE
metrics = Metrics(expected, predicted, source=file, feature=label, threshold=0.5)
results[label] = metrics.as_dict()
viz = viz_pixel_labels(expected, predicted, rgb[1:-1],
label_negative=LABEL_NEGATIVE,
label_positive=LABEL_POSITIVE)
skimage.io.imsave(os.path.join(args.output, 'file-{}-viz-{}.jpg'.format(file_index, label)), viz)
with open(cached_file, 'w') as f:
yaml.dump(results, f, Dumper=yamlordereddictloader.Dumper)
cached = yaml.load(open(cached_file))
print("Cumulative:")
for label in i12.LABELS:
if label == 'background':
continue
metrics = Metrics(**cached[label])
assert metrics.source == file
summary[label]['TP'] += metrics.TP
summary[label]['FP'] += metrics.FP
summary[label]['TN'] += metrics.TN
summary[label]['FN'] += metrics.FN
summary[label]['targets'] += metrics.targets
summary[label]['candidates'] += metrics.candidates
summary[label]['hits'] += metrics.hits
cum = Metrics(**summary[label])
print("{:10}: pix(P:{:2.5f}, R:{:2.5f},F:{:2.5f}), obj:(P:{:2.5f}, R:{:2.5f},F:{:2.5f})".format(
label,
cum.pixel_precision,
cum.pixel_recall,
cum.pixel_f1,
cum.object_precision,
cum.object_recall,
cum.object_f1))