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

TL;DR

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.

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.

Figures (9)

• Figure 1: PI3DETR is an end-to-end pipeline that takes a point cloud as input. Like 3DETR misra2021end, it uses an SAModule qi2017pointnet++ to sample points and generate queries, and a Transformer vaswani2017attention to extract query-specific features. Multiple heads predict parametric curves (cubic Bézier, line segments, circle, arc) with associated parameters, while a geometry-aware matcher aligns predictions with ground truth, removing the need for intermediate representations or post-processing. This visualization shows three sample queries for clarity.
• Figure 2: Unlike methods that rely on intermediate representations (distance fields, segmentations, voxels) and post-fitting, PI3DETR directly predicts parametric edge curves in an end-to-end manner.
• Figure 3: Parametric representations of the four curve types used for prediction: Bézier ($B_i$), Line Segment ($L_i$), Circle ($C_i$), and Arc ($A_i$).
• Figure 4: Optional post-processing steps: (a) S&F, (b) IoU Filter.
• Figure 5: Qualitative comparison of predicted curves using $N = 32{,}768$. Our method and NerVE (CAD) predict parametric curves, while NerVE (PWL) produces piece-wise linear curves. NerVE (CAD) may produce invalid predictions, shown as empty.