def run(in_json: params.input_configuration_type,
out_dir: str,
epi_step: int = 30,
region_size: int = 500,
disparity_margin: float = 0.02,
epipolar_error_upper_bound: float = 10.,
epipolar_error_maximum_bias: float = 0.,
elevation_delta_lower_bound: float = -1000.,
elevation_delta_upper_bound: float = 1000.,
mode: str = "local_dask",
nb_workers: int = 4,
walltime: str = "00:59:00",
check_inputs: bool = False):
"""
Main function of the prepare subcommand
This function will perform the following steps:
1. Compute stereo-rectification grids for the input pair
2. Compute all possible sift matches in epipolar geometry
3. Derive an optimal disparity range to explore from the matches
4. Derive a bilinear correction model of the stereo-rectification grid for right image in order to minimize epipolar error
5. Apply correction to right grid
6. Export left and corrected right grid
:param in_json: dictionary describing input data (see README.md for format)
:param out_dir: Directory where all outputs will be written, including a content.json file describing its content
:param epi_step: Step of the epipolar grid to compute (in pixels in epipolar geometry)
:param region_size: Size of regions used for sift matching
:param disparity_margin: Percent of the disparity range width to add at each end as security margin
:param epipolar_error_upper_bound: Upper bound of expected epipolar error (in pixels)
:param epipolar_error_maximum_bias: Maximum bias for epipolar error (in pixels)
:param elevation_delta_lower_bound: Lower bound for elevation delta with respect to initial MNT (in meters)
:param elevation_delta_upper_bound: Upper bound for elevation delta with respect to initial MNT (in meters)
:param mode: Parallelization mode
:param nb_workers: Number of dask workers to use for the sift matching step
:param walltime: Walltime of the dask workers
:param check_inputs: activation of the inputs consistency checking
"""
out_dir = os.path.abspath(out_dir)
# Ensure that outdir exists
try:
os.makedirs(out_dir)
except OSError as exc:
if exc.errno == errno.EEXIST and os.path.isdir(out_dir):
pass
else:
raise
utils.add_log_file(out_dir, 'prepare')
if not check_inputs:
logging.warning(
'The inputs consistency will not be checked. To enable the inputs checking, add \'--check_'
'inputs\' to your command line')
# Check configuration dict
config = utils.check_json(in_json, params.input_configuration_schema)
# Retrieve static parameters (sift and low res dsm)
static_params = static_cfg.get_cfg()
# Initialize output json dict
out_json = {
params.input_section_tag: config,
params.preprocessing_section_tag: {
params.preprocessing_version_tag: utils.get_version(),
params.preprocessing_parameters_section_tag: {
params.epi_step_tag: epi_step,
params.disparity_margin_tag: disparity_margin,
params.epipolar_error_upper_bound_tag:
epipolar_error_upper_bound,
params.epipolar_error_maximum_bias_tag:
epipolar_error_maximum_bias,
params.elevation_delta_lower_bound_tag:
elevation_delta_lower_bound,
params.elevation_delta_upper_bound_tag:
elevation_delta_upper_bound
},
params.static_params_tag: static_params[static_cfg.prepare_tag],
params.preprocessing_output_section_tag: {}
}
}
# Read input parameters
img1 = config[params.img1_tag]
img2 = config[params.img2_tag]
srtm_dir = config[params.srtm_dir_tag]
nodata1 = config.get(params.nodata1_tag, None)
nodata2 = config.get(params.nodata2_tag, None)
mask1 = config.get(params.mask1_tag, None)
mask2 = config.get(params.mask2_tag, None)
color1 = config.get(params.color1_tag, None)
if check_inputs:
logging.info('Checking inputs consistency')
if utils.rasterio_get_nb_bands(
img1) != 1 or utils.rasterio_get_nb_bands(img2) != 1:
raise Exception('{} and {} are not mono-band images'.format(
img1, img2))
if mask1 is not None:
if utils.rasterio_get_size(img1) != utils.rasterio_get_size(mask1):
raise Exception(
'The image {} and the mask {} do not have the same size'.
format(img1, mask1))
if mask2 is not None:
if utils.rasterio_get_size(img2) != utils.rasterio_get_size(mask2):
raise Exception(
'The image {} and the mask {} do not have the same size'.
format(img2, mask2))
if not utils.otb_can_open(img1):
raise Exception(
'Problem while opening image {} with the otb'.format(img1))
if not utils.otb_can_open(img2):
raise Exception(
'Problem while opening image {} with the otb'.format(img1))
with rio.open(img1) as im:
trans = im.transform
if trans.e < 0:
logging.warning(
'{} seems to have an incoherent pixel size. '
'Input images has to be in sensor geometry.'.format(img1))
with rio.open(img2) as im:
trans = im.transform
if trans.e < 0:
logging.warning(
'{} seems to have an incoherent pixel size. '
'Input images has to be in sensor geometry.'.format(img2))
# Check that the envelopes intersect one another
logging.info("Computing images envelopes and their intersection")
shp1 = os.path.join(out_dir, "left_envelope.shp")
shp2 = os.path.join(out_dir, "right_envelope.shp")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.left_envelope_tag] = shp1
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.right_envelope_tag] = shp2
preprocessing.image_envelope(img1, shp1, srtm_dir)
preprocessing.image_envelope(img2, shp2, srtm_dir)
poly1, epsg1 = utils.read_vector(shp1)
poly2, epsg2 = utils.read_vector(shp2)
inter_poly, (inter_xmin, inter_ymin, inter_xmax, inter_ymax) = \
tiling.ground_polygon_from_envelopes(poly1, poly2, epsg1, epsg2, epsg1)
out_envelopes_intersection = os.path.join(out_dir,
'envelopes_intersection.gpkg')
utils.write_vector([inter_poly], out_envelopes_intersection, epsg1)
conf_out_dict = out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag]
conf_out_dict[
params.envelopes_intersection_tag] = out_envelopes_intersection
conf_out_dict[params.envelopes_intersection_bb_tag] = [
inter_xmin, inter_ymin, inter_xmax, inter_ymax
]
if check_inputs:
logging.info('Checking DEM coverage')
dem_useful_polygon, dem_coverage = projection.compute_dem_intersection_with_poly(
srtm_dir, inter_poly, epsg1)
if dem_coverage < 100.0:
logging.warning(
'The input DEM covers {}% of the useful zone'.format(
int(dem_coverage)))
# Generate rectification grids
logging.info("Generating epipolar rectification grid ...")
grid1, grid2, epipolar_size_x, epipolar_size_y, alt_to_disp_ratio, stereogrid_pipeline = pipelines.build_stereorectification_grid_pipeline(
img1, img2, srtm_dir, epi_step)
# we want disp_to_alt_ratio = resolution/(B/H), in m.pixel^-1
disp_to_alt_ratio = 1 / alt_to_disp_ratio
# Export grids to numpy
left_grid_as_array = np.copy(
stereogrid_pipeline["stereo_app"].GetVectorImageAsNumpyArray(
"io.outleft"))
right_grid_as_array = np.copy(
stereogrid_pipeline["stereo_app"].GetVectorImageAsNumpyArray(
"io.outright"))
grid_origin = stereogrid_pipeline["stereo_app"].GetImageOrigin(
"io.outleft")
grid_spacing = stereogrid_pipeline["stereo_app"].GetImageSpacing(
"io.outleft")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.epipolar_size_x_tag] = epipolar_size_x
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.epipolar_size_y_tag] = epipolar_size_y
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.epipolar_origin_x_tag] = grid_origin[0]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.epipolar_origin_y_tag] = grid_origin[1]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.epipolar_spacing_x_tag] = grid_spacing[0]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.epipolar_spacing_y_tag] = grid_spacing[1]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.disp_to_alt_ratio_tag] = disp_to_alt_ratio
logging.info("Size of epipolar images: {}x{} pixels".format(
epipolar_size_x, epipolar_size_y))
logging.info(
"Disparity to altitude factor: {} m/pixel".format(disp_to_alt_ratio))
logging.info("Sparse matching ...")
nb_threads = int(os.environ.get('OMP_NUM_THREADS', '1'))
# Compute the full range needed for sparse matching
disp_lower_bound = elevation_delta_lower_bound / disp_to_alt_ratio
disp_upper_bound = elevation_delta_upper_bound / disp_to_alt_ratio
disparity_range_width = disp_upper_bound - disp_lower_bound
logging.info(
"Full disparity range width for sparse matching: {} pixels".format(
disparity_range_width))
disparity_range_center = (elevation_delta_upper_bound +
elevation_delta_lower_bound) / (
2 * disp_to_alt_ratio)
# Compute the number of offsets to consider in order to explore the full range
nb_splits = 1 + int(math.floor(float(disparity_range_width) / region_size))
actual_region_size = int(
math.ceil((region_size + disparity_range_width) / nb_splits))
actual_range = nb_splits * actual_region_size
actual_range_start = disparity_range_center - actual_range / 2 + region_size / 2
logging.info(
"Disparity range will be explored in {} regions of size {}, starting at {} pixels"
.format(nb_splits, actual_region_size, actual_range_start))
regions = tiling.split(0, 0, epipolar_size_x, epipolar_size_y, region_size,
region_size)
logging.info("Number of splits to process for sparse matching: {}".format(
len(regions)))
cluster = None
client = None
# TODO: prepare mp mode
# Use dask
use_dask = {"local_dask": True, "pbs_dask": True}
if mode not in use_dask.keys():
raise NotImplementedError('{} mode is not implemented'.format(mode))
if mode == "local_dask":
cluster, client = start_local_cluster(nb_workers)
else:
cluster, client = start_cluster(nb_workers, walltime, out_dir)
# Write temporary grid
tmp1 = os.path.join(out_dir, "tmp1.tif")
preprocessing.write_grid(left_grid_as_array, tmp1, grid_origin,
grid_spacing)
tmp2 = os.path.join(out_dir, "tmp2.tif")
preprocessing.write_grid(right_grid_as_array, tmp2, grid_origin,
grid_spacing)
# Compute margins for right region
margins = [
int(
math.floor(epipolar_error_upper_bound +
epipolar_error_maximum_bias)),
int(
math.floor(epipolar_error_upper_bound +
epipolar_error_maximum_bias)),
int(
math.floor(epipolar_error_upper_bound +
epipolar_error_maximum_bias)),
int(math.ceil(epipolar_error_upper_bound +
epipolar_error_maximum_bias))
]
logging.info(
"Margins added to right region for matching: {}".format(margins))
# Matching tasks as delayed objects
delayed_matches = []
for left_region in regions:
for offset in range(nb_splits):
offset_ = actual_range_start + offset * actual_region_size
# Pad region to include margins for right image
right_region = [
left_region[0] + offset_, left_region[1],
left_region[0] + offset_ + actual_region_size, left_region[3]
]
# Pad with margin and crop to largest region
right_region = tiling.crop(
tiling.pad(right_region,
margins), [0, 0, epipolar_size_x, epipolar_size_y])
# Avoid empty regions
if not tiling.empty(right_region):
delayed_matches.append(
dask.delayed(matching_wrapper)(left_region, right_region,
img1, img2, tmp1, tmp2,
mask1, mask2, nodata1,
nodata2, epipolar_size_x,
epipolar_size_y))
# Transform delayed tasks to future
logging.info("Submitting {} tasks to dask".format(len(delayed_matches)))
future_matches = client.compute(delayed_matches)
# Initialize output matches array
matches = np.empty((0, 4))
# Wait for all matching tasks to be completed
for future, result in tqdm(as_completed(future_matches, with_results=True),
total=len(future_matches),
desc="Performing matching ..."):
matches = np.concatenate((matches, result))
raw_nb_matches = matches.shape[0]
logging.info(
"Raw number of matches found: {} matches".format(raw_nb_matches))
# Export matches
logging.info("Writing raw matches file")
raw_matches_array_path = os.path.join(out_dir, "raw_matches.npy")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.raw_matches_tag] = raw_matches_array_path
np.save(raw_matches_array_path, matches)
# Filter matches that are out of margin
if epipolar_error_maximum_bias == 0:
epipolar_median_shift = 0
else:
epipolar_median_shift = np.median(matches[:, 3] - matches[:, 1])
matches = matches[((matches[:, 3] - matches[:, 1]) -
epipolar_median_shift) >= -epipolar_error_upper_bound]
matches = matches[((matches[:, 3] - matches[:, 1]) -
epipolar_median_shift) <= epipolar_error_upper_bound]
matches_discarded_message = "{} matches discarded because their epipolar error is greater \
than --epipolar_error_upper_bound = {} pix".format(
raw_nb_matches - matches.shape[0], epipolar_error_upper_bound)
if epipolar_error_maximum_bias != 0:
matches_discarded_message += " considering a shift of {} pix".format(
epipolar_median_shift)
logging.info(matches_discarded_message)
filtered_nb_matches = matches.shape[0]
matches = matches[matches[:, 2] - matches[:, 0] >= disp_lower_bound]
matches = matches[matches[:, 2] - matches[:, 0] <= disp_upper_bound]
logging.info(
"{} matches discarded because they fall outside of disparity range defined by --elevation_delta_lower_bound = {} m and --elevation_delta_upper_bound = {} m : [{} pix., {} pix.]"
.format(filtered_nb_matches - matches.shape[0],
elevation_delta_lower_bound, elevation_delta_upper_bound,
disp_lower_bound, disp_upper_bound))
# Retrieve number of matches
nb_matches = matches.shape[0]
# Check if we have enough matches
# TODO: we could also make it a warning and continue with uncorrected grid
# and default disparity range
if nb_matches < 100:
logging.critical(
"Insufficient amount of matches found (< 100), can not safely estimate epipolar error correction and disparity range"
)
# stop cluster
stop_cluster(cluster, client)
# Exit immediately
return
logging.info(
"Number of matches kept for epipolar error correction: {} matches".
format(nb_matches))
# Remove temporary files
os.remove(tmp1)
os.remove(tmp2)
# Compute epipolar error
epipolar_error = matches[:, 1] - matches[:, 3]
logging.info(
"Epipolar error before correction: mean = {:.3f} pix., standard deviation = {:.3f} pix., max = {:.3f} pix."
.format(np.mean(epipolar_error), np.std(epipolar_error),
np.max(np.fabs(epipolar_error))))
# Commpute correction for right grid
logging.info("Generating correction for right epipolar grid ...")
corrected_right_grid, corrected_matches, in_stats, out_stats = preprocessing.correct_right_grid(
matches, right_grid_as_array, grid_origin, grid_spacing)
corrected_epipolar_error = corrected_matches[:, 1] - corrected_matches[:,
3]
logging.info(
"Epipolar error after correction: mean = {:.3f} pix., standard deviation = {:.3f} pix., max = {:.3f} pix."
.format(np.mean(corrected_epipolar_error),
np.std(corrected_epipolar_error),
np.max(np.fabs(corrected_epipolar_error))))
# TODO: add stats in content.json
out_left_grid = os.path.join(out_dir, "left_epipolar_grid.tif")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.left_epipolar_grid_tag] = out_left_grid
preprocessing.write_grid(left_grid_as_array, out_left_grid, grid_origin,
grid_spacing)
# Export corrected right grid
out_right_grid = os.path.join(out_dir, "right_epipolar_grid.tif")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.right_epipolar_grid_tag] = out_right_grid
preprocessing.write_grid(corrected_right_grid, out_right_grid, grid_origin,
grid_spacing)
# Export uncorrected right grid
logging.info("Writing uncorrected right grid")
out_right_grid_uncorrected = os.path.join(
out_dir, "right_epipolar_grid_uncorrected.tif")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
right_epipolar_uncorrected_grid_tag] = out_right_grid_uncorrected
preprocessing.write_grid(right_grid_as_array, out_right_grid_uncorrected,
grid_origin, grid_spacing)
# Compute the disparity range (we filter matches that are too off epipolar
# lins after correction)
corrected_std = np.std(corrected_epipolar_error)
corrected_matches = corrected_matches[
np.fabs(corrected_epipolar_error) < 3 * corrected_std]
logging.info(
"{} matches discarded because their epipolar error is greater than 3*stdev of epipolar error after correction (3*stddev = {:.3f} pix.)"
.format(nb_matches - corrected_matches.shape[0], 3 * corrected_std))
logging.info(
"Number of matches kept for disparity range estimation: {} matches".
format(corrected_matches.shape[0]))
dmin, dmax = preprocessing.compute_disparity_range(
corrected_matches,
static_cfg.get_disparity_outliers_rejection_percent())
margin = abs(dmax - dmin) * disparity_margin
dmin -= margin
dmax += margin
logging.info(
"Disparity range with margin: [{:.3f} pix., {:.3f} pix.] (margin = {:.3f} pix.)"
.format(dmin, dmax, margin))
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.minimum_disparity_tag] = dmin
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.maximum_disparity_tag] = dmax
logging.info(
"Equivalent range in meters: [{:.3f} m, {:.3f} m] (margin = {:.3f} m)".
format(dmin * disp_to_alt_ratio, dmax * disp_to_alt_ratio,
margin * disp_to_alt_ratio))
# Export matches
logging.info("Writing matches file")
matches_array_path = os.path.join(out_dir, "matches.npy")
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.matches_tag] = matches_array_path
np.save(matches_array_path, corrected_matches)
# Now compute low resolution DSM and its initial DEM counterpart
# First, triangulate matches
logging.info("Generating low resolution DSM from matches")
points_cloud_from_matches = stereo.triangulate_matches(
out_json, corrected_matches)
# Then define the size of the lower res DSM to rasterize
low_res_dsm_params = static_cfg.get_low_res_dsm_params()
lowres_dsm_resolution = getattr(
low_res_dsm_params,
static_cfg.low_res_dsm_resolution_in_degree_tag) # Value in degree
lowres_dsm_sizex = int(
math.ceil((inter_xmax - inter_xmin) / lowres_dsm_resolution))
lowres_dsm_sizey = int(
math.ceil((inter_ymax - inter_ymin) / lowres_dsm_resolution))
lowres_dsm = rasterization.simple_rasterization_dataset(
[points_cloud_from_matches],
lowres_dsm_resolution,
4326,
color_list=None,
xstart=inter_xmin,
ystart=inter_ymax,
xsize=lowres_dsm_sizex,
ysize=lowres_dsm_sizey)
lowres_dsm_file = os.path.join(
out_dir, "lowres_dsm_from_matches.nc") # TODO add propoer crs info
lowres_dsm.to_netcdf(lowres_dsm_file)
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.lowres_dsm_tag] = lowres_dsm_file
# Now read the exact same grid on initial DEM
lowres_initial_dem = preprocessing.read_lowres_dem(
srtm_dir,
startx=inter_xmin,
starty=inter_ymax,
sizex=lowres_dsm_sizex,
sizey=lowres_dsm_sizey,
resolution=lowres_dsm_resolution)
lowres_initial_dem_file = os.path.join(out_dir, "lowres_initial_dem.nc")
lowres_initial_dem.to_netcdf(lowres_initial_dem_file)
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.lowres_initial_dem_tag] = lowres_initial_dem_file
# also write the difference
lowres_elevation_difference_file = os.path.join(
out_dir, "lowres_elevation_diff.nc")
lowres_dsm_diff = lowres_initial_dem - lowres_dsm
(lowres_dsm_diff).to_netcdf(lowres_elevation_difference_file)
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
lowres_elevation_difference_tag] = lowres_elevation_difference_file
# Now, estimate a correction to align DSM on the lowres initial DEM
splines = None
if lowres_dsm_sizex > getattr(low_res_dsm_params, static_cfg.low_res_dsm_min_sizex_for_align_tag) and \
lowres_dsm_sizey > getattr(low_res_dsm_params, static_cfg.low_res_dsm_min_sizey_for_align_tag):
logging.info(
"Estimating correction between low resolution DSM and initial DEM")
# First, we estimate direction of acquisition time for both images
vec1 = preprocessing.get_time_ground_direction(img1, dem=srtm_dir)
vec2 = preprocessing.get_time_ground_direction(img2, dem=srtm_dir)
time_direction_vector = (vec1 + vec2) / 2
display_angle = lambda x: 180 * math.atan2(x[1], x[0]) / math.pi
logging.info(
"Time direction average azimuth: {}° (img1: {}°, img2: {}°)".
format(display_angle(time_direction_vector), display_angle(vec1),
display_angle(vec2)))
origin = [
float(lowres_dsm_diff.x[0].values),
float(lowres_dsm_diff.y[0].values)
]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.time_direction_line_origin_x_tag] = origin[0]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.time_direction_line_origin_y_tag] = origin[1]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
time_direction_line_vector_x_tag] = time_direction_vector[0]
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
time_direction_line_vector_y_tag] = time_direction_vector[1]
# Then we estimate the correction splines
splines = preprocessing.lowres_initial_dem_splines_fit(
lowres_dsm,
lowres_initial_dem,
origin,
time_direction_vector,
ext=getattr(low_res_dsm_params, static_cfg.low_res_dsm_ext_tag),
order=getattr(low_res_dsm_params, static_cfg.low_res_dsm_ext_tag))
else:
logging.warning(
"Low resolution DSM is not large enough (minimum size is 100x100) to estimate correction to fit initial DEM, skipping ..."
)
if splines is not None:
# Save model to file
lowres_dem_splines_fit_file = os.path.join(
out_dir, "lowres_dem_splines_fit.pck")
with open(lowres_dem_splines_fit_file, 'wb') as f:
pickle.dump(splines, f)
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
lowres_dem_splines_fit_tag] = lowres_dem_splines_fit_file
logging.info(
"Generating corrected low resolution DSM from matches")
# Estimate correction on point cloud from matches
points_cloud_from_matches_z_correction = splines(
preprocessing.project_coordinates_on_line(
points_cloud_from_matches.x, points_cloud_from_matches.y,
origin, time_direction_vector))
# Estimate disparity correction
points_cloud_disp_correction = points_cloud_from_matches_z_correction / disp_to_alt_ratio
# Correct matches disparity
z_corrected_matches = corrected_matches
z_corrected_matches[:,
2] = z_corrected_matches[:,
2] - points_cloud_disp_correction[:,
0]
# Triangulate and rasterize again
corrected_points_cloud_from_matches = stereo.triangulate_matches(
out_json, z_corrected_matches)
corrected_lowres_dsm = rasterization.simple_rasterization_dataset(
[corrected_points_cloud_from_matches],
lowres_dsm_resolution,
corrected_points_cloud_from_matches.attrs['epsg'],
xstart=inter_xmin,
ystart=inter_ymax,
xsize=lowres_dsm_sizex,
ysize=lowres_dsm_sizey)
# Write corrected lowres dsm
corrected_lowres_dsm_file = os.path.join(
out_dir, "corrected_lowres_dsm_from_matches.nc"
) # TODO add propoer crs info
corrected_lowres_dsm.to_netcdf(corrected_lowres_dsm_file)
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
corrected_lowres_dsm_tag] = corrected_lowres_dsm_file
# also write the difference
corrected_lowres_elevation_difference_file = os.path.join(
out_dir, "corrected_lowres_elevation_diff.nc")
corrected_lowres_dsm_diff = lowres_initial_dem - corrected_lowres_dsm
(corrected_lowres_dsm_diff
).to_netcdf(corrected_lowres_elevation_difference_file)
out_json[params.preprocessing_section_tag][
params.preprocessing_output_section_tag][
params.
corrected_lowres_elevation_difference_tag] = corrected_lowres_elevation_difference_file
# Write the output json
try:
utils.check_json(out_json, params.preprocessing_content_schema)
except CheckerError as e:
logging.warning(
"content.json does not comply with schema: {}".format(e))
out_json_path = os.path.join(out_dir, "content.json")
params.write_preprocessing_content_file(out_json, out_json_path)
# stop cluster
stop_cluster(cluster, client)
def test_write_geotiff_dsm():
"""
Test to result tiles to GTiff file(s).
"""
with tempfile.TemporaryDirectory(dir=temporary_dir()) as directory:
# Set variables and parameters for create dataset
resolution = 0.5
epsg = 32630
dsm_no_data = -32768
hgt_no_data = dsm_no_data
color_no_data = 0
msk_no_data = 255
xstart, ystart, xsize, ysize = [0, 10, 10, 10]
raster = np.ndarray(shape=(10, 10, 2), dtype=np.float32)
mean = np.ndarray(shape=(10, 10, 2), dtype=np.float32)
stdev = np.ndarray(shape=(10, 10, 2), dtype=np.float32)
n_pts = np.ndarray(shape=(10, 10), dtype=np.uint16)
n_in_cell = np.ndarray(shape=(10, 10), dtype=np.uint16)
msk = np.ndarray(shape=(10, 10), dtype=np.uint16)
delayed_raster_datasets = list()
delayed_raster_datasets.append(
dask.delayed(rasterization.create_raster_dataset)(
raster, xstart, ystart, xsize, ysize, resolution, hgt_no_data,
color_no_data, epsg, mean, stdev, n_pts, n_in_cell, msk))
# Start cluster with a local cluster
nb_workers = 4
cluster, client = start_local_cluster(nb_workers)
future_dsm = client.compute(delayed_raster_datasets)
# Write netcdf with raster_datasets
bounds = [xstart, ystart, xsize, ysize]
nb_bands = 1
dsm_file = os.path.join(directory, "dsm.tif")
clr_file = os.path.join(directory, "clr.tif")
msk_file = os.path.join(directory, "msk.tif")
readwrite.write_geotiff_dsm(future_dsm,
directory,
xsize,
ysize,
bounds,
resolution,
epsg,
nb_bands,
dsm_no_data,
color_no_data,
write_msk=True,
msk_no_data=msk_no_data)
# stop cluster
stop_local_cluster(cluster, client)
width_ref = xsize
height_ref = ysize
crs = osr.SpatialReference()
crs.ImportFromEPSG(epsg)
crs_ref = crs.ExportToWkt()
geotransform = (bounds[0], resolution, 0.0, bounds[3], 0.0,
-resolution)
transform_ref = Affine.from_gdal(*geotransform)
# Compare DSM
with rio.open(dsm_file) as rio_actual:
np.testing.assert_equal(rio_actual.width, width_ref)
np.testing.assert_equal(rio_actual.height, height_ref)
assert rio_actual.transform == transform_ref
assert rio_actual.crs == crs_ref
assert rio_actual.nodata == hgt_no_data
for i in range(rio_actual.width):
for j in range(rio_actual.height):
if np.isnan(raster[i][j][0]):
assert rio_actual.read()[0][i][j] == dsm_no_data
else:
assert rio_actual.read()[0][i][j] == raster[i][j][0]
# Compare CLR
with rio.open(clr_file) as rio_actual:
np.testing.assert_equal(rio_actual.width, width_ref)
np.testing.assert_equal(rio_actual.height, height_ref)
assert rio_actual.transform == transform_ref
assert rio_actual.crs == crs_ref
assert rio_actual.nodata == color_no_data
for i in range(rio_actual.width):
for j in range(rio_actual.height):
if np.isnan(raster[i][j][1]):
assert rio_actual.read()[0][i][j] == color_no_data
else:
assert rio_actual.read()[0][i][j] == raster[i][j][1]
# Compare MSK
with rio.open(msk_file) as rio_actual:
np.testing.assert_equal(rio_actual.width, width_ref)
np.testing.assert_equal(rio_actual.height, height_ref)
assert rio_actual.transform == transform_ref
assert rio_actual.crs == crs_ref
assert rio_actual.nodata == msk_no_data
for i in range(rio_actual.width):
for j in range(rio_actual.height):
if np.isnan(raster[i][j][1]):
assert rio_actual.read()[0][i][j] == msk_no_data
else:
assert rio_actual.read()[0][i][j] == msk[i][j]
def test_local_cluster():
"""
"""
clus, client = cluster.start_local_cluster(4)
cluster.stop_local_cluster(clus, client)
def run(in_jsons: List[params.preprocessing_content_type],
out_dir: str,
resolution: float = 0.5,
min_elevation_offset: float = None,
max_elevation_offset: float = None,
epsg: int = None,
sigma: float = None,
dsm_radius: int = 1,
dsm_no_data: int = -32768,
color_no_data: int = 0,
corr_config: Dict = None,
output_stats: bool = False,
mode: str = "local_dask",
nb_workers: int = 4,
walltime: str = "00:59:00",
roi: Tuple[List[int], int] = None,
use_geoid_alt: bool = False,
use_sec_disp: bool = False,
snap_to_img1: bool = False,
align: bool = False,
cloud_small_components_filter: bool = True,
cloud_statistical_outliers_filter: bool = True,
epi_tile_size: int = None):
"""
Main function for the compute_dsm subcommand
This function will compute independent tiles of the final DSM, with the following steps:
1. Epipolar resampling (including mask)
2. Disparity map estimation
3. Triangulation of disparity map
4. Rasterization to DSM
:param in_jsons: dictionaries describing the input pair (as produced by cars_preproc tool)
:param out_dir: directory where output raster and color images will be written
:param resolution: resolution of DSM to produce
:param min_elevation_offset: Override minimum disparity from prepare step with this offset in meters
:param max_elevation_offset: Override maximum disparity from prepare step with this offset in meters
:param epsg: epsg code for the CRS of the output DSM
:param sigma: width of gaussian weight for rasterization
:param dsm_radius: Radius around a cell for gathering points for rasterization
:param dsm_no_data: No data value to use in the final DSM file
:param color_no_data: No data value to use in the final colored image
:param corr_config: Correlator configuration
:param output_stats: flag, if true, outputs dsm as a geotiff file with quality statistics.
:param mode: Parallelization mode
:param nb_workers: Number of dask workers to use for the sift matching step
:param walltime: Walltime of the dask workers
:param roi: DSM ROI in final projection with the corresponding epsg code ([xmin, ymin, xmax, ymax], roi_epsg))
(roi_epsg can be set to None if the ROI is in final projection)
:param use_geoid_alt: Wheter altitude should be computed wrt geoid height or not.
:param use_sec_disp: Boolean activating the use of the secondary disparity map
:param snap_to_img1: If this is True, Lines of Sight of img2 are moved so as to cross those of img1
:param align: If this is True, use the correction estimated during prepare to align to lowres DEM (if available)
:param cloud_small_components_filter: Boolean activating the points cloud small components filtering. The filter's
parameters are set in the static configuration json.
:param cloud_statistical_outliers_filter: Boolean activating the points cloud statistical outliers filtering.
The filter's parameters are set in the static configuration json.
:param epi_tile_size: Force the size of epipolar tiles (None by default)
"""
out_dir = os.path.abspath(out_dir)
# Ensure that outdir exists
try:
os.makedirs(out_dir)
except OSError as exc:
if exc.errno == errno.EEXIST and os.path.isdir(out_dir):
pass
else:
raise
tmp_dir = os.path.join(out_dir, 'tmp')
utils.add_log_file(out_dir, 'compute_dsm')
logging.info("Received {} stereo pairs configurations".format(
len(in_jsons)))
# Retrieve static parameters (rasterization and cloud filtering)
static_params = static_cfg.get_cfg()
# Initiate ouptut json dictionary
out_json = {
params.stereo_inputs_section_tag: [],
params.stereo_section_tag: {
params.stereo_version_tag: utils.get_version(),
params.stereo_parameters_section_tag: {
params.resolution_tag: resolution,
params.sigma_tag: sigma,
params.dsm_radius_tag: dsm_radius
},
params.static_params_tag:
static_params[static_cfg.compute_dsm_tag],
params.stereo_output_section_tag: {}
}
}
if use_geoid_alt:
geoid_data = utils.read_geoid_file()
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.alt_reference_tag] = 'geoid'
else:
geoid_data = None
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.alt_reference_tag] = 'ellipsoid'
if epsg is not None:
out_json[params.stereo_section_tag][
params.stereo_parameters_section_tag][params.epsg_tag] = epsg
roi_epsg = None
if roi is not None:
(roi_xmin, roi_ymin, roi_xmax, roi_ymax), roi_epsg = roi
roi_poly = Polygon([(roi_xmin, roi_ymin), (roi_xmax, roi_ymin),
(roi_xmax, roi_ymax), (roi_xmin, roi_ymax),
(roi_xmin, roi_ymin)])
# set the timeout for each job in multiprocessing mode (in seconds)
perJobTimeout = 600
configurations_data = {}
config_idx = 1
ref_left_image = None
for in_json in in_jsons:
# Build config id
config_id = "config_{}".format(config_idx)
# Check configuration with respect to schema
configuration = utils.check_json(in_json,
params.preprocessing_content_schema)
preprocessing_output_config = configuration[
params.preprocessing_section_tag][
params.preprocessing_output_section_tag]
# Append input configuration to output json
out_json[params.stereo_inputs_section_tag].append(configuration)
configurations_data[config_id] = {}
configurations_data[config_id]['configuration'] = configuration
# Check left image and raise a warning if different left images are used along with snap_to_img1 mpode
if ref_left_image is None:
ref_left_image = configuration[params.input_section_tag][
params.img1_tag]
else:
if snap_to_img1 and ref_left_image != configuration[
params.input_section_tag][params.img1_tag]:
logging.warning(
"--snap_to_left_image mode is used but input configurations have different images as their left image in pair. This may result in increasing registration discrepencies between pairs"
)
# Get largest epipolar regions from configuration file
largest_epipolar_region = [
0, 0, preprocessing_output_config[params.epipolar_size_x_tag],
preprocessing_output_config[params.epipolar_size_y_tag]
]
configurations_data[config_id][
'largest_epipolar_region'] = largest_epipolar_region
disp_min = preprocessing_output_config[params.minimum_disparity_tag]
disp_max = preprocessing_output_config[params.maximum_disparity_tag]
disp_to_alt_ratio = preprocessing_output_config[
params.disp_to_alt_ratio_tag]
# Check if we need to override disp_min
if min_elevation_offset is not None:
user_disp_min = min_elevation_offset / disp_to_alt_ratio
if user_disp_min > disp_min:
logging.warning((
'Overriden disparity minimum = {:.3f} pix. (or {:.3f} m.) is greater '
'than disparity minimum estimated in prepare step = {:.3f} pix. (or '
'{:.3f} m.) for configuration {}').format(
user_disp_min, min_elevation_offset, disp_min,
disp_min * disp_to_alt_ratio, config_id))
disp_min = user_disp_min
# Check if we need to override disp_max
if max_elevation_offset is not None:
user_disp_max = max_elevation_offset / disp_to_alt_ratio
if user_disp_max < disp_max:
logging.warning((
'Overriden disparity maximum = {:.3f} pix. (or {:.3f} m.) is lower '
'than disparity maximum estimated in prepare step = {:.3f} pix. (or '
'{:.3f} m.) for configuration {}').format(
user_disp_max, max_elevation_offset, disp_max,
disp_max * disp_to_alt_ratio, config_id))
disp_max = user_disp_max
logging.info(
'Disparity range for config {}: [{:.3f} pix., {:.3f} pix.] (or [{:.3f} m., {:.3f} m.])'
.format(config_id, disp_min, disp_max,
disp_min * disp_to_alt_ratio,
disp_max * disp_to_alt_ratio))
configurations_data[config_id]['disp_min'] = disp_min
configurations_data[config_id]['disp_max'] = disp_max
origin = [
preprocessing_output_config[params.epipolar_origin_x_tag],
preprocessing_output_config[params.epipolar_origin_y_tag]
]
spacing = [
preprocessing_output_config[params.epipolar_spacing_x_tag],
preprocessing_output_config[params.epipolar_spacing_y_tag]
]
configurations_data[config_id]['origin'] = origin
configurations_data[config_id]['spacing'] = spacing
logging.info(
"Size of epipolar image: {}".format(largest_epipolar_region))
logging.debug("Origin of epipolar grid: {}".format(origin))
logging.debug("Spacing of epipolar grid: {}".format(spacing))
# Warning if align is set but correction is missing
if align and params.lowres_dem_splines_fit_tag not in preprocessing_output_config:
logging.warning((
'Align with low resolution DSM option is set but splines correction file '
'is not available for configuration {}. Correction '
'will not be applied for this configuration'
).format(config_id))
# Numpy array with corners of largest epipolar region. Order
# does not matter here, since it will be passed to stereo.compute_epipolar_grid_min_max
corners = np.array(
[[[largest_epipolar_region[0], largest_epipolar_region[1]],
[largest_epipolar_region[0], largest_epipolar_region[3]]],
[[largest_epipolar_region[2], largest_epipolar_region[3]],
[largest_epipolar_region[2], largest_epipolar_region[1]]]],
dtype=np.float64)
# get utm zone with the middle point of terrain_min if epsg is
# None
if epsg is None:
# Compute terrain position of epipolar image corners for min and max disparity
terrain_dispmin, terrain_dispmax = stereo.compute_epipolar_grid_min_max(
corners, 4326, configuration, disp_min, disp_max)
epsg = rasterization.get_utm_zone_as_epsg_code(
*np.mean(terrain_dispmin, axis=0))
logging.info("EPSG code: {}".format(epsg))
# Compute terrain min and max again, this time using estimated epsg code
terrain_dispmin, terrain_dispmax = stereo.compute_epipolar_grid_min_max(
corners, epsg, configuration, disp_min, disp_max)
if roi_epsg is not None:
if roi_epsg != epsg:
roi_poly = projection.polygon_projection(
roi_poly, roi_epsg, epsg)
# Compute bounds from epipolar image corners and dispmin/dispmax
terrain_bounds = np.stack((terrain_dispmin, terrain_dispmax), axis=0)
terrain_min = np.amin(terrain_bounds, axis=(0, 1))
terrain_max = np.amax(terrain_bounds, axis=(0, 1))
terrain_area = (terrain_max[0] - terrain_min[0]) * (terrain_max[1] -
terrain_min[1])
configurations_data[config_id]['terrain_area'] = terrain_area
logging.info(
"Terrain area covered: {} square meters (or square degrees)".
format(terrain_area))
# Retrieve bounding box of the ground intersection of the envelopes
inter_poly, inter_epsg = utils.read_vector(
preprocessing_output_config[params.envelopes_intersection_tag])
if epsg != inter_epsg:
inter_poly = projection.polygon_projection(inter_poly, inter_epsg,
epsg)
(inter_xmin, inter_ymin, inter_xmax, inter_ymax) = inter_poly.bounds
# Align bounding box to integer resolution steps
xmin, ymin, xmax, ymax = tiling.snap_to_grid(inter_xmin, inter_ymin,
inter_xmax, inter_ymax,
resolution)
logging.info("Terrain bounding box : [{}, {}] x [{}, {}]".format(
xmin, xmax, ymin, ymax))
configurations_data[config_id]['terrain_bounding_box'] = [
xmin, ymin, xmax, ymax
]
if roi is not None:
if not roi_poly.intersects(inter_poly):
logging.warning(
"The pair composed of {} and {} does not intersect the requested ROI"
.format(
configuration[params.input_section_tag][
params.img1_tag], configuration[
params.input_section_tag][params.img2_tag]))
# Get optimal tile size
if epi_tile_size is not None:
opt_epipolar_tile_size = epi_tile_size
else:
opt_epipolar_tile_size = stereo.optimal_tile_size(
disp_min, disp_max)
logging.info(
"Optimal tile size for epipolar regions: {}x{} pixels".format(
opt_epipolar_tile_size, opt_epipolar_tile_size))
configurations_data[config_id][
'opt_epipolar_tile_size'] = opt_epipolar_tile_size
# Split epipolar image in pieces
epipolar_regions = tiling.split(
0, 0, preprocessing_output_config[params.epipolar_size_x_tag],
preprocessing_output_config[params.epipolar_size_y_tag],
opt_epipolar_tile_size, opt_epipolar_tile_size)
epipolar_regions_grid = tiling.grid(
0, 0, preprocessing_output_config[params.epipolar_size_x_tag],
preprocessing_output_config[params.epipolar_size_y_tag],
opt_epipolar_tile_size, opt_epipolar_tile_size)
configurations_data[config_id]['epipolar_regions'] = epipolar_regions
configurations_data[config_id][
'epipolar_regions_grid'] = epipolar_regions_grid
logging.info("Epipolar image will be processed in {} splits".format(
len(epipolar_regions)))
# Increment config index
config_idx += 1
xmin, ymin, xmax, ymax = tiling.union([
conf['terrain_bounding_box']
for config_id, conf in configurations_data.items()
])
if roi is not None:
# terrain bounding box polygon
terrain_poly = Polygon([(xmin, ymin), (xmax, ymin), (xmax, ymax),
(xmin, ymax), (xmin, ymin)])
if not roi_poly.intersects(terrain_poly):
raise Exception(
'None of the input pairs intersect the requested ROI')
else:
logging.info('Setting terrain bounding box to the requested ROI')
xmin, ymin, xmax, ymax = roi_poly.bounds
xmin, ymin, xmax, ymax = tiling.snap_to_grid(
xmin, ymin, xmax, ymax, resolution)
logging.info("Total terrain bounding box : [{}, {}] x [{}, {}]".format(
xmin, xmax, ymin, ymax))
# Compute optimal terrain tile size
optimal_terrain_tile_widths = []
for config_id, conf in configurations_data.items():
# Compute terrain area covered by a single epipolar tile
terrain_area_covered_by_epipolar_tile = conf["terrain_area"] / len(
epipolar_regions)
# Compute tile width in pixels
optimal_terrain_tile_widths.append(
math.sqrt(terrain_area_covered_by_epipolar_tile))
# In case of multiple json configuration, take the average optimal size,
# and align to multiple of resolution
optimal_terrain_tile_width = int(
math.ceil(
np.mean(optimal_terrain_tile_widths) / resolution)) * resolution
logging.info("Optimal terrain tile size: {}x{} pixels".format(
int(optimal_terrain_tile_width / resolution),
int(optimal_terrain_tile_width / resolution)))
# Split terrain bounding box in pieces
terrain_grid = tiling.grid(xmin, ymin, xmax, ymax,
optimal_terrain_tile_width,
optimal_terrain_tile_width)
number_of_terrain_splits = (terrain_grid.shape[0] -
1) * (terrain_grid.shape[1] - 1)
logging.info("Terrain bounding box will be processed in {} splits".format(
number_of_terrain_splits))
# Start dask cluster
cluster = None
client = None
# Use dask
use_dask = {"local_dask": True, "pbs_dask": True, "mp": False}
if mode not in use_dask.keys():
raise NotImplementedError('{} mode is not implemented'.format(mode))
if use_dask[mode]:
if mode == "local_dask":
cluster, client = start_local_cluster(nb_workers)
else:
cluster, client = start_cluster(nb_workers, walltime, out_dir)
# Add plugin to monitor memory of workers
plugin = ComputeDSMMemoryLogger(out_dir)
client.register_worker_plugin(plugin)
geoid_data_futures = None
if geoid_data is not None:
# Broadcast geoid data to all dask workers
geoid_data_futures = client.scatter(geoid_data, broadcast=True)
# Retrieve the epsg code which will be used for the triangulation's output points clouds
# (ecef if filters are activated)
if cloud_small_components_filter or cloud_statistical_outliers_filter:
stereo_out_epsg = 4978
else:
stereo_out_epsg = epsg
# Submit all epipolar regions to be processed as delayed tasks, and
# project terrain grid to epipolar
for config_id, conf in configurations_data.items():
# This list will hold the different epipolar tiles to be
# processed as points cloud
delayed_point_clouds = []
if use_dask[mode]:
# Use Dask delayed
for region in conf['epipolar_regions']:
delayed_point_clouds.append(
dask.delayed(stereo.images_pair_to_3d_points)(
conf['configuration'],
region,
corr_config,
disp_min=conf['disp_min'],
disp_max=conf['disp_max'],
geoid_data=geoid_data_futures,
out_epsg=stereo_out_epsg,
use_sec_disp=use_sec_disp,
snap_to_img1=snap_to_img1,
align=align))
logging.info(
"Submitted {} epipolar delayed tasks to dask for stereo configuration {}"
.format(len(delayed_point_clouds), config_id))
else:
# Use multiprocessing module
# create progress bar with an update callback
pbar = tqdm(total=len(conf['epipolar_regions']))
def update(args):
pbar.update()
# create a thread pool
pool = multiprocessing.Pool(nb_workers)
# launch several 'write_3d_points()' to process each epipolar region
for region in conf['epipolar_regions']:
delayed_point_clouds.append(
pool.apply_async(write_3d_points,
args=(conf['configuration'], region,
corr_config, tmp_dir, config_id),
kwds={
'disp_min': conf['disp_min'],
'disp_max': conf['disp_max'],
'geoid_data': geoid_data,
'out_epsg': stereo_out_epsg,
'use_sec_disp': use_sec_disp
},
callback=update))
# Wait computation results (timeout in seconds) and replace the
# async objects by the actual output of write_3d_points(), meaning
# the paths to cloud files
delayed_point_clouds = [
delayed_pc.get(timeout=perJobTimeout)
for delayed_pc in delayed_point_clouds
]
# closing thread pool when computation is done
pool.close()
pool.join()
configurations_data[config_id][
'delayed_point_clouds'] = delayed_point_clouds
# build list of epipolar region hashes
configurations_data[config_id]['epipolar_regions_hash'] = [
region_hash_string(k) for k in conf['epipolar_regions']
]
# Compute disp_min and disp_max location for epipolar grid
epipolar_grid_min, epipolar_grid_max = stereo.compute_epipolar_grid_min_max(
conf['epipolar_regions_grid'], epsg, conf["configuration"],
conf['disp_min'], conf['disp_max'])
epipolar_regions_grid_flat = conf['epipolar_regions_grid'].reshape(
-1, conf['epipolar_regions_grid'].shape[-1])
# in the following code a factor is used to increase the precision
spatial_ref = osr.SpatialReference()
spatial_ref.ImportFromEPSG(epsg)
if spatial_ref.IsGeographic():
precision_factor = 1000.0
else:
precision_factor = 1.0
# Build delaunay_triangulation
delaunay_min = Delaunay(epipolar_grid_min * precision_factor)
delaunay_max = Delaunay(epipolar_grid_max * precision_factor)
# Build kdtrees
tree_min = cKDTree(epipolar_grid_min * precision_factor)
tree_max = cKDTree(epipolar_grid_max * precision_factor)
# Look-up terrain_grid with Delaunay
s_min = tsearch(delaunay_min, terrain_grid * precision_factor)
s_max = tsearch(delaunay_max, terrain_grid * precision_factor)
points_disp_min = epipolar_regions_grid_flat[
delaunay_min.simplices[s_min]]
points_disp_max = epipolar_regions_grid_flat[
delaunay_max.simplices[s_max]]
nn_disp_min = epipolar_regions_grid_flat[tree_min.query(
terrain_grid * precision_factor)[1]]
nn_disp_max = epipolar_regions_grid_flat[tree_max.query(
terrain_grid * precision_factor)[1]]
points_disp_min_min = np.min(points_disp_min, axis=2)
points_disp_min_max = np.max(points_disp_min, axis=2)
points_disp_max_min = np.min(points_disp_max, axis=2)
points_disp_max_max = np.max(points_disp_max, axis=2)
# Use either Delaunay search or NN search if delaunay search fails (point outside triangles)
points_disp_min_min = np.where(
np.stack((s_min, s_min), axis=-1) != -1, points_disp_min_min,
nn_disp_min)
points_disp_min_max = np.where(
np.stack((s_min, s_min), axis=-1) != -1, points_disp_min_max,
nn_disp_min)
points_disp_max_min = np.where(
np.stack((s_max, s_max), axis=-1) != -1, points_disp_max_min,
nn_disp_max)
points_disp_max_max = np.where(
np.stack((s_max, s_max), axis=-1) != -1, points_disp_max_max,
nn_disp_max)
points = np.stack((points_disp_min_min, points_disp_min_max,
points_disp_max_min, points_disp_max_max),
axis=0)
points_min = np.min(points, axis=0)
points_max = np.max(points, axis=0)
configurations_data[config_id]['epipolar_points_min'] = points_min
configurations_data[config_id]['epipolar_points_max'] = points_max
# Retrieve number of bands
if params.color1_tag in configuration[params.input_section_tag]:
nb_bands = utils.rasterio_get_nb_bands(
configuration[params.input_section_tag][params.color1_tag])
else:
logging.info(
'No color image has been given in input, {} will be used as the color image'
.format(configuration[params.input_section_tag][params.img1_tag]))
nb_bands = utils.rasterio_get_nb_bands(
configuration[params.input_section_tag][params.img1_tag])
logging.info("Number of bands in color image: {}".format(nb_bands))
rank = []
# This list will contained the different raster tiles to be written by cars
delayed_dsm_tiles = []
number_of_epipolar_tiles_per_terrain_tiles = []
if not use_dask[mode]:
# create progress bar with update callback
pbar = tqdm(
total=number_of_terrain_splits,
desc="Finding correspondences between terrain and epipolar tiles")
def update(args):
pbar.update()
# initialize a thread pool for multiprocessing mode
pool = multiprocessing.Pool(nb_workers)
# Loop on terrain regions and derive dependency to epipolar regions
for terrain_region_dix in tqdm(range(number_of_terrain_splits),
total=number_of_terrain_splits,
desc="Delaunay look-up"):
j = int(terrain_region_dix / (terrain_grid.shape[1] - 1))
i = terrain_region_dix % (terrain_grid.shape[1] - 1)
logging.debug("Processing tile located at {},{} in tile grid".format(
i, j))
terrain_region = [
terrain_grid[j, i, 0], terrain_grid[j, i, 1],
terrain_grid[j + 1, i + 1, 0], terrain_grid[j + 1, i + 1, 1]
]
logging.debug(
"Corresponding terrain region: {}".format(terrain_region))
# This list will hold the required points clouds for this terrain tile
required_point_clouds = []
# For each stereo configuration
for config_id, conf in configurations_data.items():
epipolar_points_min = conf['epipolar_points_min']
epipolar_points_max = conf['epipolar_points_max']
tile_min = np.minimum(
np.minimum(
np.minimum(epipolar_points_min[j, i],
epipolar_points_min[j + 1, i]),
np.minimum(epipolar_points_min[j + 1, i + 1],
epipolar_points_min[j, i + 1])),
np.minimum(
np.minimum(epipolar_points_max[j, i],
epipolar_points_max[j + 1, i]),
np.minimum(epipolar_points_max[j + 1, i + 1],
epipolar_points_max[j, i + 1])))
tile_max = np.maximum(
np.maximum(
np.maximum(epipolar_points_min[j, i],
epipolar_points_min[j + 1, i]),
np.maximum(epipolar_points_min[j + 1, i + 1],
epipolar_points_min[j, i + 1])),
np.maximum(
np.maximum(epipolar_points_max[j, i],
epipolar_points_max[j + 1, i]),
np.maximum(epipolar_points_max[j + 1, i + 1],
epipolar_points_max[j, i + 1])))
# Bouding region of corresponding cell
epipolar_region_minx = tile_min[0]
epipolar_region_miny = tile_min[1]
epipolar_region_maxx = tile_max[0]
epipolar_region_maxy = tile_max[1]
# This mimics the previous code that was using
# transform_terrain_region_to_epipolar
epipolar_region = [
epipolar_region_minx, epipolar_region_miny,
epipolar_region_maxx, epipolar_region_maxy
]
# Crop epipolar region to largest region
epipolar_region = tiling.crop(epipolar_region,
conf['largest_epipolar_region'])
logging.debug(
"Corresponding epipolar region: {}".format(epipolar_region))
# Check if the epipolar region contains any pixels to process
if tiling.empty(epipolar_region):
logging.debug(
"Skipping terrain region because corresponding epipolar region is empty"
)
else:
# Loop on all epipolar tiles covered by epipolar region
for epipolar_tile in tiling.list_tiles(
epipolar_region, conf['largest_epipolar_region'],
conf['opt_epipolar_tile_size']):
cur_hash = region_hash_string(epipolar_tile)
# Look for corresponding hash in delayed point clouds
# dictionnary
if cur_hash in conf['epipolar_regions_hash']:
# If hash can be found, append it to the required
# clouds to compute for this terrain tile
pos = conf['epipolar_regions_hash'].index(cur_hash)
required_point_clouds.append(
conf['delayed_point_clouds'][pos])
# start and size parameters for the rasterization function
xstart, ystart, xsize, ysize = tiling.roi_to_start_and_size(
terrain_region, resolution)
# cloud filtering params
if cloud_small_components_filter:
small_cpn_filter_params = static_cfg.get_small_components_filter_params(
)
else:
small_cpn_filter_params = None
if cloud_statistical_outliers_filter:
statistical_filter_params = static_cfg.get_statistical_outliers_filter_params(
)
else:
statistical_filter_params = None
# rasterization grid division factor
rasterization_params = static_cfg.get_rasterization_params()
grid_points_division_factor = getattr(
rasterization_params, static_cfg.grid_points_division_factor_tag)
if len(required_point_clouds) > 0:
logging.debug(
"Number of clouds to process for this terrain tile: {}".format(
len(required_point_clouds)))
if use_dask[mode]:
# Delayed call to rasterization operations using all required
# point clouds
rasterized = dask.delayed(rasterization_wrapper)(
required_point_clouds,
resolution,
epsg,
xstart=xstart,
ystart=ystart,
xsize=xsize,
ysize=ysize,
radius=dsm_radius,
sigma=sigma,
dsm_no_data=dsm_no_data,
color_no_data=color_no_data,
small_cpn_filter_params=small_cpn_filter_params,
statistical_filter_params=statistical_filter_params,
grid_points_division_factor=grid_points_division_factor)
# Keep track of delayed raster tiles
delayed_dsm_tiles.append(rasterized)
rank.append(i * i + j * j)
else:
# Launch asynchrone job for write_dsm_by_tile()
delayed_dsm_tiles.append(
pool.apply_async(
write_dsm_by_tile,
args=(required_point_clouds, resolution, epsg, tmp_dir,
nb_bands, static_cfg.get_color_image_encoding(),
output_stats),
kwds={
'xstart':
xstart,
'ystart':
ystart,
'xsize':
xsize,
'ysize':
ysize,
'radius':
dsm_radius,
'sigma':
sigma,
'dsm_no_data':
dsm_no_data,
'color_no_data':
color_no_data,
'small_cpn_filter_params':
small_cpn_filter_params,
'statistical_filter_params':
statistical_filter_params,
'grid_points_division_factor':
grid_points_division_factor
},
callback=update))
number_of_epipolar_tiles_per_terrain_tiles.append(
len(required_point_clouds))
logging.info(
"Average number of epipolar tiles for each terrain tile: {}".format(
int(np.round(
np.mean(number_of_epipolar_tiles_per_terrain_tiles)))))
logging.info(
"Max number of epipolar tiles for each terrain tile: {}".format(
np.max(number_of_epipolar_tiles_per_terrain_tiles)))
bounds = (xmin, ymin, xmax, ymax)
# Derive output image files parameters to pass to rasterio
xsize, ysize = tiling.roi_to_start_and_size([xmin, ymin, xmax, ymax],
resolution)[2:]
out_dsm = os.path.join(out_dir, "dsm.tif")
out_clr = os.path.join(out_dir, "clr.tif")
out_dsm_mean = os.path.join(out_dir, "dsm_mean.tif")
out_dsm_std = os.path.join(out_dir, "dsm_std.tif")
out_dsm_n_pts = os.path.join(out_dir, "dsm_n_pts.tif")
out_dsm_points_in_cell = os.path.join(out_dir, "dsm_pts_in_cell.tif")
if use_dask[mode]:
# Sort tiles according to rank
delayed_dsm_tiles = [
delayed for _, delayed in sorted(zip(rank, delayed_dsm_tiles),
key=lambda pair: pair[0])
]
logging.info("Submitting {} tasks to dask".format(
len(delayed_dsm_tiles)))
# Transform all delayed raster tiles to futures (computation starts
# immediatly on workers, assynchronously)
future_dsm_tiles = client.compute(delayed_dsm_tiles)
logging.info("DSM output image size: {}x{} pixels".format(
xsize, ysize))
readwrite.write_geotiff_dsm(
future_dsm_tiles,
out_dir,
xsize,
ysize,
bounds,
resolution,
epsg,
nb_bands,
dsm_no_data,
color_no_data,
color_dtype=static_cfg.get_color_image_encoding(),
write_color=True,
write_stats=output_stats)
# stop cluster
stop_cluster(cluster, client)
else:
logging.info("Computing DSM tiles ...")
# Wait for asynchrone jobs (timeout in seconds) and replace them by
# write_dsm_by_tile() output
delayed_dsm_tiles = [
delayed_tile.get(timeout=perJobTimeout)
for delayed_tile in delayed_dsm_tiles
]
# closing the tread pool after computation
pool.close()
pool.join()
# vrt to tif
logging.info("Building VRT")
vrt_options = gdal.BuildVRTOptions(resampleAlg='nearest')
def vrt_mosaic(tiles_glob, vrt_name, vrt_options, output):
vrt_file = os.path.join(out_dir, vrt_name)
tiles_list = glob(os.path.join(out_dir, 'tmp', tiles_glob))
vrt = gdal.BuildVRT(vrt_file, tiles_list, options=vrt_options)
vrt = None
ds = gdal.Open(vrt_file)
ds = gdal.Translate(output, ds)
ds = None
vrt_mosaic('*_dsm.tif', 'dsm.vrt', vrt_options, out_dsm)
vrt_mosaic('*_clr.tif', 'clr.vrt', vrt_options, out_clr)
if output_stats:
vrt_mosaic('*_dsm_mean.tif', 'dsm_mean.vrt', vrt_options,
out_dsm_mean)
vrt_mosaic('*_dsm_std.tif', 'dsm_std.vrt', vrt_options,
out_dsm_std)
vrt_mosaic('*_dsm_n_pts.tif', 'dsm_n_pts.vrt', vrt_options,
out_dsm_n_pts)
vrt_mosaic('*_pts_in_cell.tif', 'dsm_pts_in_cell.vrt', vrt_options,
out_dsm_points_in_cell)
# Fill output json file
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.epsg_tag] = epsg
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.dsm_tag] = out_dsm
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.dsm_no_data_tag] = float(dsm_no_data)
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.color_no_data_tag] = float(color_no_data)
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.color_tag] = out_clr
if output_stats:
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.dsm_mean_tag] = out_dsm_mean
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.dsm_std_tag] = out_dsm_std
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.dsm_n_pts_tag] = out_dsm_n_pts
out_json[params.stereo_section_tag][params.stereo_output_section_tag][
params.dsm_points_in_cell_tag] = out_dsm_points_in_cell
# Write the output json
out_json_path = os.path.join(out_dir, "content.json")
try:
utils.check_json(out_json, params.stereo_content_schema)
except CheckerError as e:
logging.warning(
"content.json does not comply with schema: {}".format(e))
params.write_stereo_content_file(out_json, out_json_path)
def test_local_cluster():
"""
Simple start and stop local cluster test
"""
clus, client = cluster.start_local_cluster(4)
cluster.stop_local_cluster(clus, client)