def save_reconstruction_data(imp_surf_dist_ms, dataset, model_out_dir,
shape_ind):
from source import sdf
shape = dataset.shape_cache.get(shape_ind)
imp_surf_dist_ms_nan = np.isnan(imp_surf_dist_ms)
# the predicted distance would be greater than 1 -> not possible with tanh
imp_surf_dist_ms[imp_surf_dist_ms_nan] = 1.0
# save query points
os.makedirs(os.path.join(model_out_dir, 'query_pts_ms'), exist_ok=True)
np.save(
os.path.join(model_out_dir, 'query_pts_ms',
dataset.shape_names[shape_ind] + '.xyz.npy'),
shape.imp_surf_query_point_ms)
# save query distance in model space
os.makedirs(os.path.join(model_out_dir, 'dist_ms'), exist_ok=True)
np.save(
os.path.join(model_out_dir, 'dist_ms',
dataset.shape_names[shape_ind] + '.xyz.npy'),
imp_surf_dist_ms)
# debug query points with color for distance
os.makedirs(os.path.join(model_out_dir, 'query_pts_ms_vis'), exist_ok=True)
sdf.visualize_query_points(query_pts_ms=shape.imp_surf_query_point_ms,
query_dist_ms=imp_surf_dist_ms,
file_out_off=os.path.join(
model_out_dir, 'query_pts_ms_vis',
dataset.shape_names[shape_ind] + '.ply'))
def _get_and_save_query_pts(file_in_mesh: str,
file_out_query_pts: str,
file_out_query_dist: str,
file_out_query_vis: str,
num_query_pts: int,
patch_radius: float,
far_query_pts_ratio=0.1,
signed_distance_batch_size=1000,
debug=False):
import trimesh
# random state for file name
rng = np.random.RandomState(file_utils.filename_to_hash(file_in_mesh))
in_mesh = trimesh.load(file_in_mesh)
# get query pts
query_pts_ms = sdf.get_query_pts_for_mesh(in_mesh, num_query_pts,
patch_radius,
far_query_pts_ratio, rng)
np.save(file_out_query_pts, query_pts_ms.astype(np.float32))
# get signed distance
query_dist_ms = sdf.get_signed_distance(in_mesh, query_pts_ms,
signed_distance_batch_size)
# fix NaNs, Infs and truncate
nan_ids = np.isnan(query_dist_ms)
inf_ids = np.isinf(query_dist_ms)
query_dist_ms[nan_ids] = 0.0
query_dist_ms[inf_ids] = 1.0
query_dist_ms[query_dist_ms < -1.0] = -1.0
query_dist_ms[query_dist_ms > 1.0] = 1.0
np.save(file_out_query_dist, query_dist_ms.astype(np.float32))
if debug and file_out_query_vis is not None:
# save visualization
sdf.visualize_query_points(query_pts_ms, query_dist_ms,
file_out_query_vis)
def _get_and_save_query_pts(
file_in_mesh: str, file_in_query_pts: str, file_out_query_dist: str, file_out_query_vis: str,
signed_distance_batch_size=1000, debug=False):
in_mesh = trimesh.load(file_in_mesh)
# get query pts
query_pts_ms = np.load(file_in_query_pts)
# get signed distance
query_dist_ms = sdf.get_signed_distance(
in_mesh, query_pts_ms, signed_distance_batch_size)
# fix NaNs, Infs and truncate
nan_ids = np.isnan(query_dist_ms)
inf_ids = np.isinf(query_dist_ms)
query_dist_ms[nan_ids] = 0.0
query_dist_ms[inf_ids] = 1.0
query_dist_ms[query_dist_ms < -1.0] = -1.0
query_dist_ms[query_dist_ms > 1.0] = 1.0
np.save(file_out_query_dist, query_dist_ms.astype(np.float32))
if debug and file_out_query_vis is not None:
# save visualization
sdf.visualize_query_points(query_pts_ms, query_dist_ms, file_out_query_vis)
def _make_sdf_samples_from_pc(file_in_pts, file_in_normal, file_in_mesh, out_pc):
# deviation from the surface in both directions, as described in the paper, section 6.3
eta = 0.01 # actual value is never mentioned, so we just assume something small
pts = np.load(file_in_pts)
pts = pts.astype(dtype=np.float32) # DeepSDF only accepts float, not double
normals = np.loadtxt(file_in_normal)
normals_length = np.linalg.norm(normals, axis=1)
normals_length_repeated = np.repeat(np.expand_dims(normals_length, axis=1), 3, axis=1)
normals_normalized = normals / normals_length_repeated
# get sdf samples near the surface
query_pts_pos = pts + eta * normals_normalized
query_pts_neg = pts - eta * normals_normalized
signed_dist_pos = np.full((query_pts_pos.shape[0],), -eta)
signed_dist_neg = np.full((query_pts_neg.shape[0],), eta)
# combine close points and signed distance to 4D vectors
close_sdf_samples_pos = np.zeros((query_pts_pos.shape[0], 4), dtype=np.float32)
close_sdf_samples_pos[:, 0:3] = query_pts_pos
close_sdf_samples_pos[:, 3] = signed_dist_pos
close_sdf_samples_neg = np.zeros((query_pts_neg.shape[0], 4), dtype=np.float32)
close_sdf_samples_neg[:, 0:3] = query_pts_neg
close_sdf_samples_neg[:, 3] = signed_dist_neg
# get sdf samples in the unit cube (can be far away from the surface)
far_samples_ratio = 0.2
num_far_samples = int((query_pts_pos.shape[0] + query_pts_neg.shape[0]) * far_samples_ratio)
random_samples_unit_cube = np.random.rand(num_far_samples, 3) - 0.5
mesh = trimesh.load(file_in_mesh)
random_samples_unit_cube_sd = sdf.get_signed_distance(
in_mesh=mesh,
query_pts_ms=random_samples_unit_cube,
signed_distance_batch_size=1000
)
# split far sdf samples in pos and neg (inside and outside)
random_samples_unit_cube_sd_ids_pos = random_samples_unit_cube_sd > 0.0
random_samples_unit_cube_sd_ids_neg = random_samples_unit_cube_sd < 0.0
random_samples_unit_cube_pos = random_samples_unit_cube[random_samples_unit_cube_sd_ids_pos]
random_samples_unit_cube_sd_pos = random_samples_unit_cube_sd[random_samples_unit_cube_sd_ids_pos]
random_samples_unit_cube_neg = random_samples_unit_cube[random_samples_unit_cube_sd_ids_neg]
random_samples_unit_cube_sd_neg = random_samples_unit_cube_sd[random_samples_unit_cube_sd_ids_neg]
# combine far points and signed distance to 4D vectors
far_sdf_samples_pos = np.zeros((random_samples_unit_cube_pos.shape[0], 4), dtype=np.float32)
far_sdf_samples_pos[:, 0:3] = random_samples_unit_cube_pos
far_sdf_samples_pos[:, 3] = random_samples_unit_cube_sd_pos
far_sdf_samples_neg = np.zeros((random_samples_unit_cube_neg.shape[0], 4), dtype=np.float32)
far_sdf_samples_neg[:, 0:3] = random_samples_unit_cube_neg
far_sdf_samples_neg[:, 3] = random_samples_unit_cube_sd_neg
#
# # cat close and far sdf samples
# pc_sdf_pos = np.concatenate((close_sdf_samples_pos, far_sdf_samples_pos), axis=0)
# pc_sdf_neg = np.concatenate((close_sdf_samples_neg, far_sdf_samples_neg), axis=0)
#
# np.savez(out_pc, pos=pc_sdf_pos, neg=pc_sdf_neg)
# np.savez(out_pc, pos=close_sdf_samples_neg, neg=close_sdf_samples_pos)
np.savez(out_pc, pos=close_sdf_samples_neg, neg=close_sdf_samples_pos,
pos_far=far_sdf_samples_pos, neg_far=far_sdf_samples_neg)
# debug: save visualization
file_out_query_vis = out_pc + '.ply'
# query_pts_ms = np.concatenate((query_pts_pos, query_pts_neg))
query_pts_ms = np.concatenate((query_pts_pos, query_pts_neg, random_samples_unit_cube_pos, random_samples_unit_cube_neg))
# query_dist_ms = np.concatenate((signed_dist_pos, signed_dist_neg))
query_dist_ms = np.concatenate((signed_dist_pos, signed_dist_neg, random_samples_unit_cube_sd_pos, random_samples_unit_cube_sd_neg))
sdf.visualize_query_points(query_pts_ms, query_dist_ms, file_out_query_vis)
print('wrote vis to {}'.format(file_out_query_vis))
def visualize_result(pts_query_ms, dist_query_ms):
out_vis_file = os.path.join(model_out_dir, 'vis',
dataset.shape_names[shape_ind] + '.ply')
file_utils.make_dir_for_file(out_vis_file)
from source import sdf
sdf.visualize_query_points(pts_query_ms, dist_query_ms, out_vis_file)