def down_sample(
dense_pcd_path, dense_label_path, sparse_pcd_path, sparse_label_path, voxel_size
):
# Skip if done
if os.path.isfile(sparse_pcd_path) and (
not os.path.isfile(dense_label_path) or os.path.isfile(sparse_label_path)
):
print("Skipped:", file_prefix)
return
else:
print("Processing:", file_prefix)
# Inputs
dense_pcd = open3d.io.read_point_cloud(dense_pcd_path)
try:
dense_labels = load_labels(dense_label_path)
except:
dense_labels = None
# Skip label 0, we use explicit frees to reduce memory usage
print("Num points:", np.asarray(dense_pcd.points).shape[0])
if dense_labels is not None:
non_zero_indexes = dense_labels != 0
dense_points = np.asarray(dense_pcd.points)[non_zero_indexes]
dense_pcd.points = open3d.utility.Vector3dVector()
dense_pcd.points = open3d.utility.Vector3dVector(dense_points)
del dense_points
dense_colors = np.asarray(dense_pcd.colors)[non_zero_indexes]
dense_pcd.colors = open3d.utility.Vector3dVector()
dense_pcd.colors = open3d.utility.Vector3dVector(dense_colors)
del dense_colors
dense_labels = dense_labels[non_zero_indexes]
print("Num points after 0-skip:", np.asarray(dense_pcd.points).shape[0])
# Downsample points
min_bound = dense_pcd.get_min_bound() - voxel_size * 0.5
max_bound = dense_pcd.get_max_bound() + voxel_size * 0.5
sparse_pcd, cubics_ids = open3d.geometry.PointCloud.voxel_down_sample_and_trace(
dense_pcd, voxel_size, min_bound, max_bound, False
)
print("Num points after down sampling:", np.asarray(sparse_pcd.points).shape[0])
open3d.io.write_point_cloud(sparse_pcd_path, sparse_pcd)
print("Point cloud written to:", sparse_pcd_path)
# Downsample labels
if dense_labels is not None:
sparse_labels = []
for cubic_ids in cubics_ids:
cubic_ids = cubic_ids[cubic_ids != -1]
cubic_labels = dense_labels[cubic_ids]
sparse_labels.append(np.bincount(cubic_labels).argmax())
sparse_labels = np.array(sparse_labels)
write_labels(sparse_label_path, sparse_labels)
print("Labels written to:", sparse_label_path)
try:
dense_gt_labels = load_labels(os.path.join(gt_dir, file_prefix + ".labels"))
print("dense_gt_labels loaded", flush=True)
except:
print("dense_gt_labels not found, treat as test set")
dense_gt_labels = None
# Assign labels
start = time.time()
dense_labels, dense_colors = interpolator.interpolate_labels(
sparse_points, sparse_labels, dense_points
)
print("KNN interpolation time: ", time.time() - start, "seconds", flush=True)
# Write dense labels
write_labels(dense_labels_path, dense_labels)
print("Dense labels written to:", dense_labels_path, flush=True)
# Write dense point cloud with color
dense_pcd.colors = open3d.utility.Vector3dVector(dense_colors)
open3d.io.write_point_cloud(dense_points_colored_path, dense_pcd)
print("Dense pcd with color written to:", dense_points_colored_path, flush=True)
# Eval
if dense_gt_labels is not None:
cm = ConfusionMatrix(9)
cm.increment_from_list(dense_gt_labels, dense_labels)
cm.print_metrics()
cm_global.increment_from_list(dense_gt_labels, dense_labels)
pprint("Global results")