def __init__(self, data_path):
"""
Create dataset instance. Empty directories (for EXIF, matches, etc) will be created if they don't exist
already.
:param data_path: Path to directory containing dataset
"""
self.data_path = data_path
# Load configuration.
config_file = os.path.join(self.data_path, 'config.yaml')
self.config = config.load_config(config_file)
# Load list of images.
image_list_file = os.path.join(self.data_path, 'image_list.txt')
if os.path.isfile(image_list_file):
with open(image_list_file) as fin:
lines = fin.read().splitlines()
self.set_image_list(lines)
else:
self.set_image_path(os.path.join(self.data_path, 'images'))
# Create output folders.
for p in [self.__exif_path(),
self.__feature_path(),
self.__matches_path()]:
io.mkdir_p(p)
def export(self, reconstruction, graph, data):
exporter = csfm.OpenMVSExporter()
for camera in reconstruction.cameras.values():
if camera.projection_type == 'perspective':
w, h = camera.width, camera.height
K = np.array([
[camera.focal, 0, (w - 1.0) / 2 / max(w, h)],
[0, camera.focal, (h - 1.0) / 2 / max(w, h)],
[0, 0, 1],
])
exporter.add_camera(str(camera.id), K)
for shot in reconstruction.shots.values():
if shot.camera.projection_type == 'perspective':
image_path = data._undistorted_image_file(shot.id)
exporter.add_shot(
str(os.path.abspath(image_path)),
str(shot.id),
str(shot.camera.id),
shot.pose.get_rotation_matrix(),
shot.pose.get_origin())
for point in reconstruction.points.values():
shots = graph[point.id].keys()
coordinates = np.array(point.coordinates, dtype=np.float64)
exporter.add_point(coordinates, shots)
io.mkdir_p(data.data_path + '/openmvs')
exporter.export(data.data_path + '/openmvs/scene.mvs')
def create_submodels(self, clusters, no_symlinks=False):
data = DataSet(self.data_path)
for i, cluster in enumerate(clusters):
# create sub model dirs
submodel_path = self._submodel_path(i)
submodel_images_path = self._submodel_images_path(i)
io.mkdir_p(submodel_path)
io.mkdir_p(submodel_images_path)
# link images and create image list file
image_list_path = os.path.join(submodel_path, 'image_list.txt')
with io.open_wt(image_list_path) as txtfile:
for image in cluster:
src = data.image_files[image]
dst = os.path.join(submodel_images_path, image)
if not os.path.isfile(dst):
os.symlink(src, dst)
dst_relpath = os.path.relpath(dst, submodel_path)
txtfile.write(dst_relpath + "\n")
# copy config.yaml if exists
config_file_path = os.path.join(self.data_path, 'config.yaml')
if os.path.exists(config_file_path):
shutil.copyfile(config_file_path, os.path.join(submodel_path, 'config.yaml'))
if no_symlinks:
reference_file_path = os.path.join(self.data_path, 'reference_lla.json')
if os.path.exists(reference_file_path):
shutil.copyfile(reference_file_path, os.path.join(submodel_path, 'reference_lla.json'))
else:
# create symlinks to metadata files
for symlink_path in ['camera_models.json', 'reference_lla.json',
'exif', 'features', 'matches']:
self._create_symlink(submodel_path, symlink_path)
def create_submodels(self, clusters):
data = DataSet(self.data_path)
for i, cluster in enumerate(clusters):
# create sub model dirs
submodel_path = self._submodel_path(i)
submodel_images_path = self._submodel_images_path(i)
io.mkdir_p(submodel_path)
io.mkdir_p(submodel_images_path)
# create image list file
image_list_path = os.path.join(submodel_path, 'image_list.txt')
with io.open_wt(image_list_path) as txtfile:
for image in cluster:
src = data.image_files[image]
dst = os.path.join(submodel_images_path, image)
src_relpath = os.path.relpath(src, submodel_images_path)
if not os.path.isfile(dst):
os.symlink(src_relpath, dst)
dst_relpath = os.path.relpath(dst, submodel_path)
txtfile.write(dst_relpath + "\n")
# copy config.yaml if exists
config_file_path = os.path.join(self.data_path, 'config.yaml')
if os.path.exists(config_file_path):
shutil.copyfile(config_file_path, os.path.join(submodel_path, 'config.yaml'))
# create symlinks to additional files
filenames = ['camera_models.json', 'reference_lla.json', 'exif',
'features', 'matches', 'masks', 'mask_list.txt',
'segmentations']
for filename in filenames:
self._create_symlink(submodel_path, filename)
def extract_keyframes_from_video(video, reconstruction):
'''
Reads video and extracts a frame for each shot in reconstruction
'''
image_path = 'images'
mkdir_p(image_path)
T = 0.1 # TODO(pau) get this from config
cap = cv2.VideoCapture(video)
video_idx = 0
shot_ids = sorted(reconstruction['shots'].keys())
for shot_id in shot_ids:
shot = reconstruction['shots'][shot_id]
timestamp = shot['created_at']
keyframe_idx = int(round(timestamp / T))
while video_idx <= keyframe_idx:
for i in range(20):
ret, frame = cap.read()
if ret:
break
else:
print 'retrying'
if not ret:
raise RuntimeError(
'Cound not find keyframe {} in video'.format(shot_id))
if video_idx == keyframe_idx:
cv2.imwrite(os.path.join(image_path, shot_id), frame)
video_idx += 1
cap.release()
def save_pruned_depthmap(self, image, points, normals, colors, labels, detections):
io.mkdir_p(self._depthmap_path())
filepath = self._depthmap_file(image, 'pruned.npz')
np.savez_compressed(filepath,
points=points, normals=normals,
colors=colors, labels=labels,
detections=detections)
def import_video_with_gpx(video_file, gpx_file, output_path, dx, dt=None, start_time=None, visual=False, image_description=None):
points = geotag_from_gpx.get_lat_lon_time(gpx_file)
orientation = video_orientation(video_file)
if start_time:
video_start_time = dateutil.parser.parse(start_time)
else:
try:
exifdate = Popen(['exiftool', '-CreateDate', '-b', video_file], stdout=PIPE).stdout.read()
video_start_time = datetime.datetime.strptime(exifdate,'%Y:%m:%d %H:%M:%S')
except:
print 'Video recording timestamp not found. Using first GPS point time.'
video_start_time = points[0][0]
try:
duration = Popen(['exiftool', '-MediaDuration', '-b', video_file], stdout=PIPE).stdout.read()
video_duration = float(duration)
video_end_time = video_start_time + datetime.timedelta(seconds=video_duration)
except:
print 'Video end time not found. Using last GPS point time.'
video_end_time = points[-1][0]
print 'GPS track starts at:', points[0][0]
print 'Video starts at:', video_start_time
# Extract video frames.
io.mkdir_p(output_path)
key_points = geotag_from_gpx.sample_gpx(points, dx, dt)
cap = cv2.VideoCapture(video_file)
image_files = []
for p in key_points:
dt = (p[0] - video_start_time).total_seconds()
if dt > 0:
CAP_PROP_POS_MSEC = cv2.CAP_PROP_POS_MSEC if context.OPENCV3 else cv2.cv.CV_CAP_PROP_POS_MSEC
cap.set(CAP_PROP_POS_MSEC, int(dt * 1000))
ret, frame = cap.read()
if ret:
print 'Grabbing frame for time', p[0]
filepath = os.path.join(output_path, p[0].strftime("%Y_%m_%d_%H_%M_%S_%f")[:-3] + '.jpg')
cv2.imwrite(filepath, frame)
geotag_from_gpx.add_exif_using_timestamp(filepath, points, timestamp=p[0], orientation=orientation)
# Display the resulting frame
if visual:
# Display the resulting frame
max_display_size = 800
resize_ratio = float(max_display_size) / max(frame.shape[0], frame.shape[1])
frame = cv2.resize(frame, dsize=(0, 0), fx=resize_ratio, fy=resize_ratio)
cv2.imshow('frame', frame)
if cv2.waitKey(1) & 0xFF == 27:
break
image_files.append(filepath)
# When everything done, release the capture
cap.release()
if visual:
cv2.destroyAllWindows()
return image_files
def __save_features(self, filepath, image, points, descriptors, colors=None):
io.mkdir_p(self.__feature_path())
feature_type = self.config.get('feature_type')
if ((feature_type == 'AKAZE' and self.config.get('akaze_descriptor') in ['MLDB_UPRIGHT', 'MLDB']) or
(feature_type == 'HAHOG' and self.config.get('hahog_normalize_to_uchar', False))):
feature_data_type = np.uint8
else:
feature_data_type = np.float32
np.savez(filepath,
points=points.astype(np.float32),
descriptors=descriptors.astype(feature_data_type),
colors=colors)
def __init__(self, data_path):
'''
Create meta dataset instance for large scale reconstruction.
:param data_path: Path to directory containing meta dataset
'''
self.data_path = data_path
config_file = os.path.join(self.data_path, 'config.yaml')
self.config = config.load_config(config_file)
self._image_list_file_name = 'image_list_with_gps.tsv'
self._clusters_file_name = 'clusters.npz'
self._clusters_with_neighbors_file_name = 'clusters_with_neighbors.npz'
self._clusters_with_neighbors_geojson_file_name = 'clusters_with_neighbors.geojson'
io.mkdir_p(self._submodels_path())
def save_exif(self, image, data):
io.mkdir_p(self._exif_path())
with io.open_wt(self._exif_file(image)) as fout:
io.json_dump(data, fout)
def save_epipolar(self, im1, im2, R, t, X=[], inliers=[]):
io.mkdir_p(self.__epipolar_path())
np.savez(self.__epipolar_file(im1, im2), R=R, t=t, X=X, inliers=inliers)
def save_undistorted_image(self, image, array):
io.mkdir_p(self._undistorted_image_path())
cv2.imwrite(self._undistorted_image_file(image), array[:, :, ::-1])
def save_report(self, report_str, path):
"""Save report string to a file."""
filepath = os.path.join(self._report_path(), path)
io.mkdir_p(os.path.dirname(filepath))
with io.open_wt(filepath) as fout:
return fout.write(report_str)
def save_raw_depthmap(self, image, depth, plane, score, nghbr, nghbrs):
io.mkdir_p(self._depthmap_path())
filepath = self._depthmap_file(image, 'raw.npz')
np.savez_compressed(filepath, depth=depth, plane=plane, score=score, nghbr=nghbr, nghbrs=nghbrs)
def save_undistorted_segmentation(self, image, array):
"""Save the undistorted image segmentation."""
io.mkdir_p(self._undistorted_segmentation_path())
io.imwrite(self._undistorted_segmentation_file(image), array)
def save_clean_depthmap(self, image, depth, plane, score):
io.mkdir_p(self._depthmap_path())
filepath = self._depthmap_file(image, 'clean.npz')
np.savez_compressed(filepath, depth=depth, plane=plane, score=score)
def save_raw_depthmap(self, image, depth, plane, score):
io.mkdir_p(self._depthmap_path())
filepath = self._depthmap_file(image, 'raw.npz')
np.savez(filepath, depth=depth, plane=plane, score=score)
def save_undistorted_shot_ids(self, ushot_dict):
filename = os.path.join(self.data_path, "undistorted_shot_ids.json")
io.mkdir_p(self.data_path)
with io.open_wt(filename) as fout:
io.json_dump(ushot_dict, fout, minify=False)
def save_exif(self, image, data):
io.mkdir_p(self.__exif_path())
with open(self.__exif_file(image), 'w') as fout:
fout.write(io.json_dumps(data))
def save_matches(self, image, matches):
io.mkdir_p(self.__matches_path())
with gzip.open(self.__matches_file(image), 'wb') as fout:
pickle.dump(matches, fout)
def save_undistorted_reconstruction(self, reconstruction):
io.mkdir_p(self.data_path)
return self.base.save_reconstruction(reconstruction,
filename=os.path.join(
self.subfolder,
'reconstruction.json'))
def save_clean_depthmap(self, image, depth, plane, score):
io.mkdir_p(self._depthmap_path())
filepath = self._depthmap_file(image, 'clean.npz')
np.savez_compressed(filepath, depth=depth, plane=plane, score=score)
def save_undistorted_segmentation(self, image, array):
"""Save the undistorted image segmentation."""
io.mkdir_p(self._undistorted_segmentation_path())
io.imwrite(self._undistorted_segmentation_file(image), array)
def save_undistorted_mask(self, image, array):
"""Save the undistorted image mask."""
io.mkdir_p(self._undistorted_mask_path())
io.imwrite(self._undistorted_mask_file(image), array)
def save_undistorted_image(self, image, array):
"""Save undistorted image pixels."""
io.mkdir_p(self._undistorted_image_path())
io.imwrite(self._undistorted_image_file(image), array)
'submodels, images are grouped into clusters. '
'This value regulates the number of images that '
'each cluster should have on average.')
parser.add_argument('--submodel-overlap',
type=float,
metavar='<positive integer>',
default=150,
help='Radius of the overlap between submodels. '
'After grouping images into clusters, images '
'that are closer than this radius to a cluster '
'are added to the cluster. This is done to ensure '
'that neighboring submodels overlap.')
return parser.parse_args()
if __name__ == '__main__':
args = parse_command_line()
data_path = args.dataset
resize_images(data_path, args)
image_path = os.path.join(data_path, 'images')
opensfm_path = os.path.join(data_path, 'opensfm')
mkdir_p(opensfm_path)
create_image_list(image_path, opensfm_path)
create_config(opensfm_path, args)
link_image_groups(data_path, opensfm_path)
def save_report(self, report_str, path):
"""Save report string to a file."""
filepath = os.path.join(self._report_path(), path)
io.mkdir_p(os.path.dirname(filepath))
with io.open_wt(filepath) as fout:
return fout.write(report_str)
def save_undistorted_mask(self, image, array):
"""Save the undistorted image mask."""
io.mkdir_p(self._undistorted_mask_path())
io.imwrite(self._undistorted_mask_file(image), array)
def save_epipolar(self, im1, im2, R, t, X=[], inliers=[]):
io.mkdir_p(self.__epipolar_path())
np.savez(self.__epipolar_file(im1, im2), R=R, t=t, X=X, inliers=inliers)
def save_raw_depthmap(self, image, depth, plane, score, nghbr, nghbrs):
io.mkdir_p(self._depthmap_path())
filepath = self._depthmap_file(image, 'raw.npz')
np.savez_compressed(filepath, depth=depth, plane=plane, score=score, nghbr=nghbr, nghbrs=nghbrs)
def save_undistorted_image(self, image, array):
io.mkdir_p(self._undistorted_image_path())
cv2.imwrite(self._undistorted_image_file(image), array[:, :, ::-1])
def save_exif(self, image, data):
io.mkdir_p(self._exif_path())
with io.open_wt(self._exif_file(image)) as fout:
io.json_dump(data, fout)
def _save_features(self, filepath, points, descriptors, colors=None):
io.mkdir_p(self._feature_path())
features.save_features(filepath, points, descriptors, colors,
self.config)
def save_matches(self, image, matches):
io.mkdir_p(self._matches_path())
with gzip.open(self._matches_file(image), 'wb') as fout:
pickle.dump(matches, fout)
def save_exif(self, image, data):
io.mkdir_p(self.__exif_path())
with open(self.__exif_file(image), 'w') as fout:
io.json_dump(data, fout)
def save_undistorted_image(self, image, array):
"""Save undistorted image pixels."""
io.mkdir_p(self._undistorted_image_path())
io.imwrite(self._undistorted_image_file(image), array)
def import_video_with_gpx(video_file,
gpx_file,
output_path,
dx,
dt=None,
start_time=None,
visual=False,
image_description=None):
points = geotag_from_gpx.get_lat_lon_time(gpx_file)
orientation = video_orientation(video_file)
if start_time:
video_start_time = dateutil.parser.parse(start_time)
else:
try:
exifdate = Popen(['exiftool', '-CreateDate', '-b', video_file],
stdout=PIPE).stdout.read()
video_start_time = datetime.datetime.strptime(
exifdate, '%Y:%m:%d %H:%M:%S')
except:
print(
'Video recording timestamp not found. Using first GPS point time.'
)
video_start_time = points[0][0]
try:
duration = Popen(['exiftool', '-MediaDuration', '-b', video_file],
stdout=PIPE).stdout.read()
video_duration = float(duration)
video_end_time = video_start_time + datetime.timedelta(
seconds=video_duration)
except:
print('Video end time not found. Using last GPS point time.')
video_end_time = points[-1][0]
print('GPS track starts at: {}'.format(points[0][0]))
print('Video starts at: {}'.format(video_start_time))
# Extract video frames.
io.mkdir_p(output_path)
key_points = geotag_from_gpx.sample_gpx(points, dx, dt)
cap = cv2.VideoCapture(video_file)
image_files = []
for p in key_points:
dt = (p[0] - video_start_time).total_seconds()
if dt > 0:
CAP_PROP_POS_MSEC = cv2.CAP_PROP_POS_MSEC if context.OPENCV3 else cv2.cv.CV_CAP_PROP_POS_MSEC
cap.set(CAP_PROP_POS_MSEC, int(dt * 1000))
ret, frame = cap.read()
if ret:
print('Grabbing frame for time {}'.format(p[0]))
filepath = os.path.join(
output_path,
p[0].strftime("%Y_%m_%d_%H_%M_%S_%f")[:-3] + '.jpg')
cv2.imwrite(filepath, frame)
geotag_from_gpx.add_exif_using_timestamp(
filepath, points, timestamp=p[0], orientation=orientation)
# Display the resulting frame
if visual:
# Display the resulting frame
max_display_size = 800
resize_ratio = float(max_display_size) / max(
frame.shape[0], frame.shape[1])
frame = cv2.resize(frame,
dsize=(0, 0),
fx=resize_ratio,
fy=resize_ratio)
cv2.imshow('frame', frame)
if cv2.waitKey(1) & 0xFF == 27:
break
image_files.append(filepath)
# When everything done, release the capture
cap.release()
if visual:
cv2.destroyAllWindows()
return image_files
