R3PM-Net Datasets

This repository contains the two proposed datasets in R3PM-Net paper; Sioux-Cranfield and Sioux-Scans, which aim to address the gap between synthetic datasets and real-world industrial data. It also contains the pickle files made from a subset of the Sioux-Cranfield dataset that can be used to train models.

Folder Structure

R3PM-Net/
├── README.md
├── down_sampled_modelnet40.zip
├── simulators.zip
├── sioux_cranfield.zip
└── sioux_scans.zip

Downsampled ModelNet40

To save time, we provide a downsampled version of ModelNet40 test set. All the point clouds are downsampled to 2000 points.

Simulators

This directory contains pickle (.pkl) files compatible with the Learning3d library and can be used to train or fine-tune models. These files are created from a subset of the Sioux-Cranfield containing the "teeth", "cube", "lime" and "lego" CAD models. There are 320 point cloud pairs in total, with 80-20 train-test split.

Sioux-Cranfield

This is a diverse collection of 13 objects designed to evaluate model robustness across varying data qualities. The dataset contains 4 computer-aided design (CAD) models generated via photogrammetric reconstruction, 3 synthetic CAD models, and 6 pristine geometries from the Cranfield Benchmark. This combination allows for a comprehensive evaluation of performance on both high-quality synthetic meshes and realistically imperfect reconstructions.

Composition of the Sioux-Cranfield Dataset

This table provides a structured breakdown of the composition of this dataset.

Category	Source Type	Qty
Sioux (Reconstructed)	Photogram.	4
Sioux (Synthetic)	CAD Models	3
Cranfield [1]	Pristine	6
Total	---	13

Sioux (Reconstructed) and Sioux (Synthetic) are also used in Sioux-Scans dataset to produce Target point clouds.

CAD models of the Sioux-Cranfield dataset.The first six belong to the Cranfield Assembly benchmark and the rest are contributions of this paper (Sioux dataset).

Sioux-Scans

This dataset addresses the real-world challenge of registering physical scans to digital models. The targets are CAD models of seven small objects (shared with Sioux-Cranfield), while the sources are raw event-camera scans of the corresponding objects acquired via the custom Quality Control Sioux 3DoP setup. To generate these scans, the setup utilizes a laser beam and an event-based camera to produce accurate point clouds from moving or handheld objects. Unlike traditional frame-based sensors, this camera captures discrete brightness changes as the laser sweeps across the surface, resulting in highly precise point clouds. Before processing, gross outliers were filtered. However, these data represent a substantially more challenging setting than synthetic benchmarks, as they reflect inevitable deficiencies, such as sparsity, noise, and occlusions, rarely present in ideal simulated datasets. These artifacts stem from sensor noise, lighting sensitivity, and viewpoint-dependent gaps, particularly on sharp edges or reflective surfaces.

Sioux-Scans point cloud data. Target (blue) and Source (yellow) point clouds for seven distinct objects.