PI3DETR: Parametric Instance Detection of 3D Point Cloud Edges with a Geometry-Aware 3DETR

ArXiv 2025
Fabio F. Oberweger*, Michael Schwingshackl*, Vanessa Staderini,
* Equal contribution
AIT Austrian Institute of Technology
Paper arXiv Code Data
Architecture Preview

PI3DETR can directly predict 3D parametric curve instances from raw point clouds. No intermediate representations or post-processing steps are required.

Abstract

We present PI3DETR, an end-to-end framework that directly predicts 3D parametric curve instances from raw point clouds, avoiding the intermediate representations and multi-stage processing common in prior work. Extending 3DETR, our model introduces a geometry-aware matching strategy and specialized loss functions that enable unified detection of differently parameterized curve types, including cubic Bézier curves, line segments, circles, and arcs, in a single forward pass. Optional post-processing steps further refine predictions without adding complexity. This streamlined design improves robustness to noise and varying sampling densities, addressing critical challenges in real world LiDAR and 3D sensing scenarios. PI3DETR sets a new state-of-the-art on the ABC dataset and generalizes effectively to real sensor data, offering a simple yet powerful solution for 3D edge and curve estimation.

3D Parametric Edge Predictions

Explore our 3D parametric edge predictions on an unseen test set curated from the ABC dataset . The examples below showcase our model’s ability to infer precise geometric primitives directly from raw point clouds. The evaluation was conducted on our released test dataset .

At first glance, ground truth and prediction might seem identical. Zoom in and rotate the point cloud to appreciate the model’s precision 😉

Bézier Curves
Line Segments
Circles
Arcs

Colors represent different parametric edge types detected by PI3DETR from raw point clouds.

Interactive Edge Prompting

As a byproduct of our method based on a non-parametric query design, where encoded points are used as transformer decoder queries, we can construct an interactive edge generator driven by point-based prompting. The example below demonstrates this functionality on a sample from our unseen test dataset. This can be used to interactively measure distances or angles between edges, or to create custom edge combinations for CAD applications.

Real-World 3D Scan

Beyond synthetic point clouds sampled from ABC meshes, our model also operates effectively on real-world 3D scans. The examples below were obtained from 3D-printed models that were scanned using a 7-DOF robotic system equipped with a structured-light sensor. The interactive visualization highlights this capability, and additional real-world results are presented in the paper.

BibTeX

@misc{oberweger2025pi3detrparametricinstancedetection,
      title={PI3DETR: Parametric Instance Detection of 3D Point Cloud Edges with a Geometry-Aware 3DETR}, 
      author={Fabio F. Oberweger and Michael Schwingshackl and Vanessa Staderini},
      year={2025},
      eprint={2509.03262},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.03262}, 
}

This webpage was build by using the template: Nerfies.

This webpage was build by using the template: Nerfies.