pcrobust

Companion package to Do We Still Need Rotation-Invariant Architectures? A Statistical Robustness Benchmark for Point Cloud Classification (WACV 2027).

It packages the paper's benchmark as reusable parts:

• Corruption generators: SO(3) pose, Gaussian noise, and hidden-point-removal occlusion, with the paper's exact five-step severity ladders, as pure functions over (N, 3) arrays.

• Models: the 10 benchmark architectures (5 standard, 3 rotation-invariant, 2 rotation-equivariant), one Embedder API.

• Evaluation protocol: embedding-drift and accuracy sweeps, ModelNet40-C, the fine-grained rotation sweep, and the head-independent probe.

• Statistical pipeline: pilot, power analysis, confirmatory test, exploratory ANOVA, bootstrap CIs, difference-in-differences, TOST.

• Training recipes: the baseline recipe and the invariance-regularized (invreg) recipe, including the warmup and pose270 variants.

• Frozen paper data: the table-level result CSVs behind the paper (123 files), readable offline with no GPU.

• Paper reproduction (Tier 1): every table and figure of the paper (27 tables, 4 figures) rebuilt from the frozen data in seconds, and checked number by number against what the paper prints.

• Model zoo: all 193 trained checkpoints behind the paper, hosted as assets of the v1.0.0 GitHub release and downloaded on first use.

Install

pip install pcrobust

From a clone, for development:

pip install -e ".[figures,dev]"

Python 3.10–3.13. Extras: figures (matplotlib), download (gdown, for the ShapeNet archive), dev (pytest, ruff).

Quickstart

import numpy as np
import pcrobust
from pcrobust import paper

pts = np.random.rand(1024, 3).astype(np.float32)
occluded = pcrobust.corrupt(pts, "occlusion", severity=3)   # paper ladder, step 3

net = pcrobust.get_model("dgcnn", num_classes=10)           # untrained
pcrobust.list_models()                                       # all 10 keys

df = paper.load("summary/pose_primary_test.csv")             # frozen table data

pcrobust list models
pcrobust corrupt cloud.ply --corruption occlusion --severity 3
pcrobust paper list
pcrobust paper show pose_primary_test
pcrobust reproduce --list
pcrobust reproduce tab:noise --compare-paper
pcrobust reproduce all --out repro/ --compare-paper
pcrobust prepare modelnet40 --download
pcrobust evaluate robustness --arch pointnet --perturbation pose
pcrobust train invreg --arch riconvpp --device cuda
pcrobust stats --help

pcrobust evaluate|train|stats <command> --help lists each command's options. Each command is also runnable as python -m <module>.

Reproducing the paper

pcrobust reproduce rebuilds a table or figure from the bundled frozen data, offline and without a GPU. Targets are the paper's LaTeX labels (tab:noise, fig:rotation_sweep) or their bare names (noise). --out DIR writes each target as CSV, plus a PNG for figures (needs the figures extra). --compare-paper checks every rebuilt number against the number printed in the paper. A number matches when the printed value is a correct rounding of it. The command exits with status 1 if any cell differs.

from pcrobust.paper import reproduce

df = reproduce.build("tab:m40c")         # laid out like the printed table
rep = reproduce.compare("tab:m40c")      # rep.mismatches, rep.uncovered
reproduce.plot("fig:teaser", "teaser.png")

The printed numbers live in src/pcrobust/paper/expected/. That is a snapshot of the paper source, keyed to its git commit. After the paper changes, refresh it with:

python tools/sync_paper.py PATH/TO/PAPER/REPO

Two targets are partial because some of their inputs are not bundled: the latency column of tab:supp-cost, and the aug-only column of tab:supp-w4-types. pcrobust reproduce --list names them.

Data and checkpoints

Everything downloaded or produced lives under PCROBUST_HOME (default ~/.cache/pcrobust):

<PCROBUST_HOME>/
  <domain>/_raw_download/                 raw archives
  <domain>/data/{train,test,gallery}/     prepared point clouds
  checkpoints/wacv2027/<domain>/<key>.pth checkpoints (zoo download target)
  results/...                             evaluation and statistics outputs

A checkpoint key is an architecture plus an optional recipe suffix: pointnet, pointnet_invreg, pointnet_invreg_warmup, pointnet_invreg_pose270. Every evaluator accepts a directory override (--models_dir / --ckpt_root) for checkpoints kept elsewhere.

Released checkpoints download on first use and are verified against the SHA256 in src/pcrobust/models/zoo_manifest.json:

model, classes = pcrobust.load_model("modelnet40/riconvpp_invreg_rep1.pth")

pcrobust.models.zoo.list_checkpoints() lists all 193 names.

ModelNet40 downloads automatically (pcrobust prepare modelnet40 --download). ABO fetches its meshes during pcrobust prepare abo once listings.json.gz and 3dmodels.csv.gz from the public ABO bucket are in its _raw_download/. ShapeNet, ScanObjectNN, and S3DIS are registration- or request-gated: place the archive under <PCROBUST_HOME>/<domain>/_raw_download/ and run pcrobust prepare <domain>.

Tests

pytest tests

Covers every module importing standalone, a CPU forward pass for all 10 architectures, checkpoint loading, corruption properties and determinism, every frozen data file, and the CLI. The golden tests rebuild every paper table and figure and compare each printed number.

License

MIT for code. Checkpoints and result CSVs: CC BY 4.0.