if __name__ == '__main__':
# parse arguments:
arguments = get_arguments()
# load point cloud:
point_cloud = read_modelnet40_normal(arguments.input)
# compute surface normals:
# point_cloud.estimate_normals(
# search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1, max_nn=30)
# )
# build search tree:
search_tree = o3d.geometry.KDTreeFlann(point_cloud)
# detect keypoints:
keypoints = detect(point_cloud, search_tree, arguments.radius)
# visualize:
# paint background as grey:
point_cloud.paint_uniform_color([0.50, 0.50, 0.50])
# show roi:
max_bound = point_cloud.get_max_bound()
min_bound = point_cloud.get_min_bound()
center = (min_bound + max_bound) / 2.0
min_bound[1] = max_bound[1] - 0.1
max_bound[1] = max_bound[1]
min_bound[2] = center[2]
max_bound[2] = max_bound[2]
def main(input_dir, radius, bins, num_evaluations):
"""
Run point cloud registration on Shenlan dataset
"""
registration_results = io.read_registration_results(
os.path.join(input_dir, 'reg_result.txt'))
# init output
df_output = io.init_output()
for i, r in progressbar.progressbar(list(registration_results.iterrows())):
# for interactive visualization:
if i >= num_evaluations:
exit(0)
# parse point cloud index:
idx_target = int(r['idx1'])
idx_source = int(r['idx2'])
# load point clouds:
pcd_source = io.read_point_cloud_bin(
os.path.join(input_dir, 'point_clouds', f'{idx_source}.bin'))
pcd_source, idx_inliers = pcd_source.remove_radius_outlier(
nb_points=4, radius=radius)
search_tree_source = o3d.geometry.KDTreeFlann(pcd_source)
pcd_target = io.read_point_cloud_bin(
os.path.join(input_dir, 'point_clouds', f'{idx_target}.bin'))
pcd_target, idx_inliers = pcd_target.remove_radius_outlier(
nb_points=4, radius=radius)
search_tree_target = o3d.geometry.KDTreeFlann(pcd_target)
# detect keypoints:
keypoints_source = detect(pcd_source, search_tree_source, radius)
keypoints_target = detect(pcd_target, search_tree_target, radius)
# create descriptions:
pcd_source_keypoints = pcd_source.select_by_index(
keypoints_source['id'].values)
fpfh_source_keypoints = o3d.registration.compute_fpfh_feature(
pcd_source_keypoints,
o3d.geometry.KDTreeSearchParamHybrid(radius=5 * radius,
max_nn=100)).data
pcd_target_keypoints = pcd_target.select_by_index(
keypoints_target['id'].values)
fpfh_target_keypoints = o3d.registration.compute_fpfh_feature(
pcd_target_keypoints,
o3d.geometry.KDTreeSearchParamHybrid(radius=5 * radius,
max_nn=100)).data
# generate matches:
distance_threshold_init = 1.5 * radius
distance_threshold_final = 1.0 * radius
# RANSAC for initial estimation:
init_result = ransac_match(
pcd_source_keypoints,
pcd_target_keypoints,
fpfh_source_keypoints,
fpfh_target_keypoints,
ransac_params=RANSACParams(
max_workers=5,
num_samples=4,
max_correspondence_distance=distance_threshold_init,
max_iteration=200000,
max_validation=500,
max_refinement=30),
checker_params=CheckerParams(
max_correspondence_distance=distance_threshold_init,
max_edge_length_ratio=0.9,
normal_angle_threshold=None))
# exact ICP for refined estimation:
final_result = exact_match(pcd_source, pcd_target, search_tree_target,
init_result.transformation,
distance_threshold_final, 60)
# visualize:
visualize.show_registration_result(pcd_source_keypoints,
pcd_target_keypoints,
init_result.correspondence_set,
pcd_source, pcd_target,
final_result.transformation)
# add result:
io.add_to_output(df_output, idx_target, idx_source,
final_result.transformation)
# write output:
io.write_output(os.path.join(input_dir, 'reg_result_yaogefad.txt'),
df_output)
def main(input_dir, radius, bins, num_evaluations):
"""
Run pose estimation on given point cloud pair
"""
# load source & target point clouds:
pcd_source = o3d.io.read_point_cloud(os.path.join(input_dir, "first.pcd"))
pcd_source = pcd_source.voxel_down_sample(voxel_size=0.05)
pcd_source.estimate_normals(
search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1,
max_nn=30))
pcd_target = o3d.io.read_point_cloud(os.path.join(input_dir, "second.pcd"))
pcd_target = pcd_target.voxel_down_sample(voxel_size=0.05)
pcd_target.estimate_normals(
search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1,
max_nn=30))
# build search trees:
pcd_source, idx_inliers = pcd_source.remove_radius_outlier(nb_points=4,
radius=radius)
search_tree_source = o3d.geometry.KDTreeFlann(pcd_source)
pcd_target, idx_inliers = pcd_target.remove_radius_outlier(nb_points=4,
radius=radius)
search_tree_target = o3d.geometry.KDTreeFlann(pcd_target)
# detect keypoints:
keypoints_source = detect(pcd_source, search_tree_source, radius)
keypoints_target = detect(pcd_target, search_tree_target, radius)
# create descriptions:
pcd_source_keypoints = pcd_source.select_by_index(
keypoints_source['id'].values)
fpfh_source_keypoints = o3d.registration.compute_fpfh_feature(
pcd_source_keypoints,
o3d.geometry.KDTreeSearchParamHybrid(radius=5 * radius,
max_nn=100)).data
pcd_target_keypoints = pcd_target.select_by_index(
keypoints_target['id'].values)
fpfh_target_keypoints = o3d.registration.compute_fpfh_feature(
pcd_target_keypoints,
o3d.geometry.KDTreeSearchParamHybrid(radius=5 * radius,
max_nn=100)).data
# generate matches:
distance_threshold_init = 1.5 * radius
distance_threshold_final = 1.0 * radius
# RANSAC for initial estimation:
init_result = ransac_match(
pcd_source_keypoints,
pcd_target_keypoints,
fpfh_source_keypoints,
fpfh_target_keypoints,
ransac_params=RANSACParams(
max_workers=5,
num_samples=4,
max_correspondence_distance=distance_threshold_init,
max_iteration=200000,
max_validation=500,
max_refinement=30),
checker_params=CheckerParams(
max_correspondence_distance=distance_threshold_init,
max_edge_length_ratio=0.9,
normal_angle_threshold=None))
# exact ICP for refined estimation:
final_result = exact_match(pcd_source, pcd_target, search_tree_target,
init_result.transformation,
distance_threshold_final, 60)
# visualize:
visualize.show_registration_result(pcd_source_keypoints,
pcd_target_keypoints,
init_result.correspondence_set,
pcd_source, pcd_target,
final_result.transformation)
# init output
df_output = io.init_output()
# add result:
io.add_to_output(df_output, 2, 1, final_result.transformation)
# write output:
io.write_output(os.path.join(input_dir, 'reg_result_yaogefad.txt'),
df_output)