title: Examples gallery — 3D algorithms and models on public datasets description: Twelve executed examples, each a self-contained script on a bundled dataset with its image and numbers committed: registration, ICP comparison, label map to mesh, Poisson, threshold segmentation, LiDAR clustering, atlas transfer, Gaussian splats, metrics, PointNet, JEPA, VAE, DINOv2.

Examples gallery¶

Each example is one folder under examples/ with a self-contained run.py that downloads what it needs through pl.load(...), runs, prints a JSON summary and writes its images next to itself. Every image and number here was produced by the script beside it on 2026-10-16.

pip install "pardesline[all]"      # PyTorch examples: pip install torch (CPU is enough)
python examples/<folder>/run.py

Example	Dataset	What it shows	Result
register_ct_mr_rire	`rire_brain`	SimpleITK rigid registration, CT fixed, MR moving; the README demo	21 s cold; rotation [0.34, −0.35, −4.38]°
icp_registration_bunny	`open3d_bunny`	textbook Open3D ICP vs `target.align(moving)` on a 35°, 70 %-overlap, noisy copy	32.0° vs 0.14° residual
hippocampus_mask_to_mesh	`msd_task04_hippocampus`	label map to exact surfaces, Taubin smoothing, STL + sidecar	two meshes over the MRI slices
poisson_reconstruction_bunny	`open3d_bunny`	20 000 points, outlier removal, Poisson at the defaults	clean density-trimmed surface
spleen_ct_threshold_segmentation	`msd_task09_spleen`	soft-tissue window + connected components vs the expert label	the baseline a model must beat
kitti_lidar_ground_and_objects	`kitti_raw_2011_09_26_drive_0001`	raw `(N, 4)` binary through the layout, ground RANSAC, DBSCAN	35 476 points clustered
atlas_label_transfer_bunny	`open3d_bunny`	labelled atlas registered onto a noisy rotated copy, labels transferred	99.4 % agreement, 0.06°
points_to_gaussians_bunny	`open3d_bunny`	50 000 points to a 3DGS PLY	12 MB file, opens in gsplat
metrics_liver_interobserver	`liver_tumor_patient01`	Dice, HD95, surface Dice between three radiologists	Dice 0.82–0.85, HD95 2.5 mm
pointnet_hippocampus	`msd_task04_hippocampus`	40 cases to labelled point clouds, PointNet trained 40 epochs on CPU	0.97 mIoU, 0.98 accuracy
jepa_hippocampus	`msd_task04_hippocampus`	3D JEPA pretraining, laptop-sized	loss 0.48 → 0.26 in 3 epochs
vae_anomaly_hippocampus	`msd_task04_hippocampus`	VAE on 24 normal slices, 8 scored, 2 with a synthetic lesion	lesions 0.59 / 0.64, normals ≤ 0.33
dinov2_features_kitti	`kitti_raw_2011_09_26_drive_0001`	frozen DINOv2-small patch features, PCA-coloured	10×32 grid, 384-d

Images¶

CT slice and registered MR overlay SimpleITK rigid registration on RIRE

Two bunny point clouds aligned RANSAC + ICP on the Stanford bunny, 0.14° off

Hippocampus meshes over MRI slices Label map to mesh on MSD hippocampus

Bunny points and the Poisson surface Poisson reconstruction on the Stanford bunny

Threshold segmentation vs reference spleen label Threshold segmentation on MSD spleen

KITTI sweep with ground removed and objects coloured Ground removal and clustering on KITTI

Bunny atlas in grey with the scan coloured by transferred labels Atlas label transfer on the Stanford bunny

Bunny Gaussian centres coloured by height Gaussian splats on the Stanford bunny

Test slices, VAE reconstructions and residual maps with scores VAE anomaly detection on MSD hippocampus slices

KITTI camera frame above its PCA-coloured DINOv2 patch features DINOv2 features on KITTI

Real cases¶

examples/real_cases/ is where runs on data that looks like production data go (industrial scans, de-identified medical volumes, field LiDAR), with the template and the review rules. The README's "Validated on real cases" table only lists folders that exist there or in examples/.

Writing one¶

Copy templates/new_pipeline/: a run.py that loads a bundled dataset, runs one algorithm or model, writes one image and a JSON summary, and a README stating the dataset, the licence, the runtime and what the defaults handled.