Using pardesline: where everything is¶

Using pardesline¶

pardesline (import as pl) is one Python package for point clouds, meshes, medical volumes and label maps. Everything under pl.load / pl.run executes on the machine that runs Python; nothing leaves it. The cloud client is optional and separate (pardesline.cloud).

Install (tell the user exactly this)¶

pip install "pardesline[all]"       # Open3D + PyVista + SimpleITK + huggingface_hub
pip install "pardesline[local]"     # point clouds, meshes, SDF, LiDAR (Open3D, PyVista)
pip install "pardesline[medical]"   # volumes, DICOM, registration, segmentation (SimpleITK + PyVista)
pip install "pardesline[hf]"        # pl.load("hf://...")
pip install pardesline              # catalog + object model + cloud client only, no native libs

Python 3.10 to 3.12 for [local] (Open3D 0.19 has no 3.13 wheels). A missing library raises pl.MissingDependency whose message names the extra to install.

The three entry points¶

import pardesline as pl

obj = pl.load("scan.ply")  # Object (PointCloud / Mesh / Volume / LabelMap ...)
out = obj.run("point_cloud_processing.voxel_downsample", voxel_size=0.02)  # new Object
res = pl.run("point_cloud_processing", "voxel_downsample", "scan.ply", voxel_size=0.02)  # RunResult
ds = pl.load("rire_brain")  # Dataset: samples with typed roles, fetched on access

You want to	Use	Skill
run one of the 73 algorithms, see its parameters	`pl.run`, `pl.describe("module.method")`, CLI `pardesline run`	catalog-and-run
chain operations on a file, keep frame/licence/provenance	`pl.load(file)` then `.downsample() .poisson() .decimate() ...`	object-model
segment a volume, turn a label into a surface, register two scans	`Volume.segment()`, `LabelMap.to_mesh(label)`, `Volume.align(moving)`	object-model, medical-volumes
load a public dataset, a Hugging Face repo, a bucket, a TCIA collection or your own folder	`pl.load("msd_task02_heart")`, `pl.load("hf://org/name/sub")`, `pl.load("tcia://Pancreas-CT")`, `pl.load("s3://openneuro.org/ds000001")`, `pl.load("/data/study")`	datasets-and-manifests, scanning-private-folders
read a raw `.npy` / `.bin` array, standardise a dataset	`pl.load("sweep.bin", layout={...})`, `ds.export(folder)`, `pardesline export`	object-model, datasets-and-manifests
write a manifest, answer the scan's questions	`pl.scan(folder)`, `pl.load(folder, license=..., units=..., labels=...)`	scanning-private-folders
share a result, run on the hosted GPU, open a browser viewer	`pl.login()`, `pl.upload`, `pl.share`, `pardesline.cloud`	cloud-client
give an AI agent (Claude, Gemini, Cursor, Codex) the datasets and algorithms through MCP	`pardesline mcp`, `.mcp.json`, tools `load_sample`, `scan_dataset`, `run_method`	mcp-server
choose a 3D representation, convert between them	conversion matrix, memory, pitfalls	3d-representations
align two clouds or an atlas and judge the result	RANSAC + FPFH + ICP parameters, metrics, CPD	point-cloud-registration, atlas-label-transfer
register CT with MRI (or PET, CBCT)	Mattes MI, initialiser, pyramid, TRE	multimodal-rigid-registration
train or use a segmentation model on volumes or points	nnU-Net export, TotalSegmentator, PointNet data contract	medical-segmentation-models, pointnet-segmentation
use foundation-model features, pretrain without labels, detect anomalies	frozen DINOv3 heads, JEPA, VAE residual maps	dinov3-dense-features, jepa-pretraining, vae-anomaly-detection
write or read Gaussian splats	points -> 3DGS, PLY layout, training	gaussian-splatting
report a result honestly	Dice / HD95 / NSD / chamfer / TRE / AUROC, per-case statistics	evaluation-metrics-3d
add an algorithm or a manifest to the package	`src/pardesline/ops/<module>/`, `catalog/data/<module>.json`, `data/manifests/`	contributing-algorithms

Rules the package enforces (do not fight them, explain them)¶

Never guess silently. A format that carries no unit or axis convention (PLY, STL, OBJ, PCD, VTK grids) gives obj.units is None. Declare with obj.with_frame("Z-up", "mm") or a manifest. Volumes read LPS/mm from their header; glTF is Y-up/m by specification.
Parameters are validated before anything runs against the catalog (type, range, choices, unknown names with a "did you mean"). Read the error: it lists what is accepted.
Two-input methods need moving= (registration). Without it the call raises instead of self-aligning a synthetic copy.
A folder needs a confirmed manifest. pl.load(folder) raises NeedsConfirmation listing licence / units / labels questions; answer them in the call, in the saved draft, interactively, or pass assume=True.
Licences travel with the data. ds.license.describe(); bundled public datasets whose terms are not established say TO CONFIRM.

Errors (all derive from `pl.PardesLineError`, each has `.code`)¶

ParameterError (bad or unknown parameter), CatalogError (unknown module/method), RunError (script failed: .returncode, .log, .command), MissingDependency (install an extra), ManifestError, FetchError (download or checksum), NeedsConfirmation (.questions, .draft_path). FileNotFoundError for missing inputs.

Quick facts¶

19 modules, 73 methods, 371 parameters (pl.catalog()), 43 bundled datasets (pl.datasets(): 35 medical - DICOM, NIfTI, MetaImage, npz, STL, splats - over https, Hugging Face, remote archives by HTTP Range, S3, TCIA and URL indexes), outputs default to pl.output_dir() (system temp, change with pl.set_output_dir), downloads cache in ~/.cache/pardesline (PARDESLINE_CACHE). pl.o3d, pl.pv, pl.sitk, pl.np are the real libraries when installed, the cloud-executed mirrors otherwise.

_{Source: skills/using-pardesline/SKILL.md, one of the twenty skills served by pardesline mcp; published on 2026-10-14.}