Our proposed BRepDetNet detects BRep boundaries and junctions by minimizing focal-loss and
non-maximal suppression (NMS) during training time. Our main contributions are:
BRep annotations on detailed industrial-level 3D scans from the CC3D and ABC datasets.
A neural network that learns to detect BRep boundaries and junctions from a 3D scan.
Our detection network reduces false positives and negatives directly during training, achieving
notable improvements in boundary and junction detection recall compared to methods that use NMS
as a post-prediction.
Network Architecture
Our neural network is comprised of seperate Boundary Detection Heads and
Junction Detection Heads. Both network heads use DGCNN
as point feature encoder that maps \(D: S \to \phi \in \mathbb{R}^{N \times 128}\) resulting into 128
dimensional point-level deep feature vectors. The two DGCNN
encoders output boundary and junction embedding vectors \(\phi_B\) and \(\phi_J\) respectively. Next, we apply
fully-connected layer to resize fc(\(\phi_B\)) and fc(\(\phi_J\)) with final dimensions in \(\mathbb{R}^{N \times 1}\).
Dataset and Annotations
We have meticulously annotated around 50K CC3D scans and 45K High-resolution Meshes with topological
relations (e.g., next, mate, previous) between geometrical primitives (boundaries, junctions, loops, faces) in their BRep data structures. We invite the
community to use these annotations. You can access the dataset via this link: Dataset ( !Coming Soon ).
Qualitative Analysis
Visual results for boundary and junction prediction using ComplexGen/PieNet/BRepDetNet. As can be seen in
the renderer, our model gives impressive results when compared to other models. (The red points are the boundary
points and the green points are the junction points)
Quantitative Analysis
Recall and Precision of BRepDetNet
Quantitiative Results on boundary and junction prediction tasks on ABC and CC3D dataset.
Note: All the models are trained on ABC dataset
Ablation Study on Cross-Dataset Training and Evlauation
Quantitative results of an ablation study for BRepDetNet with or without NMS loss. Recall and precision for
Boundary detection on ABC
and CC3D dataset
are reported. We also showcase cross-dataset generalization ability of our model.
Video
Citation
If you refer to the results or codes of this work, please cite the following:
@inproceedings{Khan_2024_CVPR,
title = {CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention},
author = {Khan, Mohammad Sadil and Dupont, Elona and Ali, Sk Aziz and Cherenkova, Kseniya and Kacem, Anis and Aouada, Djamila},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2024}
}
@inproceedings{ali2024brepdetnet,
author = {Ali, Sk Aziz and Khan, Mohammad Sadil and Stricker, Didier},
booktitle = {2024 IEEE 3rd International Conference on Computing and Machine Intelligence (ICMI)},
title = {BRep Boundary and Junction Detection for CAD Reverse Engineering},
year = {2024},
keywords={Training;Solid modeling;Three-dimensional displays;Reverse engineering;Neural networks;Machining;Mechanical systems;BRep;Boundary Detection;Junction Detection;Scan-to-CAD;Reverse Engineering;NMS},
doi = {10.1109/ICMI60790.2024.10585950}
}
Acknowledgement
This work was partially funded by the EU Horizon Europe Framework Program under grant agreement
101058236 (HumanTech).
Authors thank Prof. Djamila Aouada and Dr. Anis Kacem (Snt, University of Luxembourg) for their valuable
inputs in understanding Scan-to-BRep paradigm.
