CstNet: Constraint-Aware Feature Learning for Parametric Point Cloud¶

Conference: ICCV 2025 arXiv: 2411.07747 Code: https://cstnetwork.github.io/ Area: 3D Vision / Point Cloud Analysis Keywords: CAD point cloud, constraint representation, parametric shape, point cloud classification, rotation robustness

TL;DR¶

This paper proposes CstNet, the first constraint-aware feature learning method for parametric point clouds. CAD constraints are encoded as point-level MAD-Adj-PT triplet representations, and a two-stage network (constraint acquisition + constraint feature learning) achieves state-of-the-art results on the newly constructed Param20K dataset, with classification accuracy improved by +3.49% and rotation robustness improved by +26.17%.

Background & Motivation¶

Parametric point clouds are sampled from CAD models and are widely used in industrial manufacturing. Existing deep learning methods tailored for CAD data suffer from three core issues:

Focus on geometry while ignoring constraints: Most methods (e.g., ParSeNet, HPNet) focus on the regularity of primitives (planes, cylinders, etc.) but overlook the constraint relationships between primitives (coaxiality, parallelism, distance, etc.). This prevents them from distinguishing CAD parts that appear similar but serve different functions — for example, eccentric wheels and flywheels look alike but have entirely different constraints (eccentric vs. coaxial).

Reliance on specialized labels: Many methods require additional annotations such as primitive types, parameters, and surface normals, whereas real-world point clouds typically only provide coordinate information.

Challenges in constraint learning: Constraints describe relationships between pairs of primitives; how to encode these into representations suitable for deep learning remains an open problem.

Method¶

Overall Architecture¶

CstNet consists of two stages: - Stage 1 (Constraint Acquisition): Extracts or predicts MAD-Adj-PT constraint representations from BRep data (CstBRep) or raw point clouds (CstPnt). - Stage 2 (Constraint Feature Learning): Fuses geometric features and constraint features via attention mechanisms to perform downstream classification.

Key Designs¶

MAD-Adj-PT Constraint Representation (Deep Learning-Friendly Constraint Encoding):
- Traditional CAD constraints are converted into three point-level components:
  - MAD (Main Axis Direction): The principal axis direction of the primitive to which a point belongs (plane → normal vector; cylinder/cone → rotation axis).
  - Adj (Adjacency): Whether a point lies near an edge (one-hot encoded), indicating connectivity between primitives.
  - PT (Primitive Type): The type of primitive a point belongs to (plane/cylinder/cone, one-hot encoded).
- The complete representation per point is \((x, y, z, \text{MAD}, \text{Adj}, \text{PT})\).
- Design Motivation: Transforming pairwise primitive relationships into relationships between each primitive and a shared reference significantly reduces data volume; the MAD+Adj combination can encode most common constraints such as parallelism, perpendicularity, and distance.
CstPnt Module (Predicting Constraints from Raw Point Clouds):
- Simulates the four-step manual constraint extraction pipeline: search for neighboring points → identify points on the same primitive → fit the shape → compute MAD/Adj/PT.
- Proposes SurfaceKNN: neighborhood search along the shape surface (rather than conventional spherical KNN), better suited to the characteristics of CAD shapes.
- Applies Q-K-V attention to assign different weights to neighboring points: \(\mathbf{f}'_i = \sum_{f_j \in \mathcal{N}_i} \rho(\text{MLP}(\text{Q}(\mathbf{f}_i) - \text{K}(\mathbf{f}_j))) \odot \text{V}(\mathbf{f}_i)\)
- Uses only local features, enabling generalization to unseen datasets after pretraining on the ABC dataset.
- Design Motivation: Removes dependency on BRep data; the limited variety of primitive types in CAD shapes (plane + cylinder + cone covers ~94%) allows local patterns to transfer across datasets.
Stage 2 Constraint Feature Learning Network:
- Triple MLP: Concatenates xyz with MAD, Adj, and PT separately to produce Axis Feature, Adjacency Feature, and Primitive Feature.
- C-MLP: Concatenates xyz with the full MAD-Adj-PT to produce an initial Constraint Feature.
- Quartic SSA: Four parallel branches process the four feature types independently, each applying FPS + SurfaceKNN + point-level attention.
- Fea Attention: Feature-level attention that adaptively reweights different features — points near edges attend more to Adjacency Feature, while other points may weight Primitive Feature more heavily.
- Design Motivation: Different constraint components carry different importance for different regions of a point cloud, necessitating adaptive weighting.

Loss & Training¶

Negative Log Likelihood Loss is used.
Adam optimizer, initial lr = 0.0001, StepLR (step = 20, gamma = 0.7), weight decay = 0.0001.
Training for 200 epochs with batch size = 16.
Stage 1 is pretrained on 25 trunks (~250K BRep files) of the ABC dataset; weights are frozen before being applied in Stage 2.

Key Experimental Results¶

Main Results¶

Param20K Classification Results:

Method	Acc(%)	Acc*(%)	F1	mAP(%)
PointNet	81.30	83.21	82.06	85.18
PointNet++	83.70	86.37	85.30	87.94
DGCNN	85.40	87.28	86.43	89.17
PTMamba	86.45	87.28	86.63	91.34
CstNet	89.94	91.06	90.34	92.72

Rotation robustness: CstNet exhibits the smallest accuracy drop when the training set is unchanged and the test set is rotated along +Z (outperforming the previous SOTA by 26.17%).

Constraint Prediction Accuracy (ABC → Param20K Generalization):

Method	MAD(MSE)↓	Adj(%)↑	PT(%)↑
ParSeNet	0.2247	81.42	59.75
HPNet	0.2570	78.60	57.66
CstNet	0.1390	87.95	86.52

Ablation Study¶

Contribution of Each Constraint Component (SurfaceKNN):

MAD	Adj	PT	Acc(%)	mAP(%)
✓	✓	✓	89.94	92.72
✗	✓	✓	86.32	90.59
✓	✗	✓	88.14	91.35
✓	✓	✗	88.68	91.22
✗	✗	✗	83.99	86.15

Stage 2 Backbone Comparison (Using Predicted Constraints):

Backbone	Acc(%)	mAP(%)
PointNet	83.83	87.27
PointNet++	85.52	90.79
DGCNN	86.81	91.35
CstNet Stage 2	89.94	92.72

Key Findings¶

Among the three constraint components, MAD contributes the most (removing it reduces accuracy by 3.62%).
SurfaceKNN outperforms standard KNN in most cases, primarily by enhancing the effectiveness of Primitive Feature.
CstPnt exhibits strong rotation robustness (dashed curves are unaffected by rotation) because the ABC training set contains local features from diverse orientations.
Using predicted constraints, while inferior to ground-truth constraint labels, still significantly outperforms the constraint-free baseline (89.94% vs. 83.99%).
Validation experiments on prisms vs. rectangular boxes intuitively demonstrate that geometric methods fail to distinguish the two when angles approach 90°, whereas the constraint-aware method maintains high accuracy throughout the 50°–87° range.

Highlights & Insights¶

First constraint-aware CAD point cloud analysis framework, filling a gap in the field.
The MAD-Adj-PT representation is elegantly designed: complex pairwise primitive relationships are simplified into point-level attributes, substantially reducing computational complexity.
Emphasis on practicality: CstPnt requires only raw point cloud input, generalizes across datasets, and does not depend on BRep data.
The newly constructed Param20K dataset fills the gap in benchmarks for parametric CAD classification.
The 26.17% improvement in rotation robustness is a particularly notable result.

Limitations & Future Work¶

Only three primitive types are considered (plane/cylinder/cone); constraint representations for free-form surfaces require further investigation.
The Param20K dataset is limited in scale (75 categories, ~20K instances), and industrial scenarios may be considerably more complex.
Computational overhead analysis of SurfaceKNN is absent.
Other downstream CAD tasks such as segmentation and assembly have not been explored.
The robustness of CstPnt on very small patches or noisy point clouds has not been validated.

ParSeNet and HPNet are representative methods for CAD point cloud analysis but focus solely on primitive fitting.
BRep-based methods such as BRepNet operate on topological data but require specialized data formats.
The core insight of this paper: function determines constraints, and constraints determine shape — parts that appear similar but serve different functions should be distinguished from a constraint perspective.
The MAD-Adj-PT framework is extensible to tasks such as CAD retrieval and assembly prediction.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ — First to introduce constraints into point cloud deep learning; representation design is original.
Experimental Thoroughness: ⭐⭐⭐⭐ — Multi-task evaluation and comprehensive ablations; dataset scale and number of baselines could be expanded.
Writing Quality: ⭐⭐⭐⭐ — Motivation is clear; validation experiments proceed in a well-structured manner.
Value: ⭐⭐⭐⭐ — Strong prospects for industrial applications; opens a new research direction.