SpelsNet: Surface Primitive Elements Segmentation by B-Rep Graph Structure Supervision

Authors: Kseniya Cherenkova, Elona Dupont, Anis Kacem, Gleb Gusev, Djamila Aouada

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To train Spels Net with the proposed point-to-BRep adjacency supervision, we extend two existing CAD datasets with the required annotations, and conduct a thorough experimental validation on them. The obtained results showcase the efficacy of Spels Net and its topological supervision compared to a set of baselines and state-of-the-art approaches.
Researcher Affiliation Collaboration Kseniya Cherenkova Sn T, University of Luxembourg, Artec3D kseniya.cherenkova@uni.lu Elona Dupont Sn T, University of Luxembourg elona.dupont@uni.lu Anis Kacem Sn T, University of Luxembourg anis.kacem@uni.lu Gleb Gusev Artec3D gleb@artec3d.com Djamila Aouada Sn T, University of Luxembourg djamila.auoada@uni.lu
Pseudocode No The paper describes the network architecture and components in text and with a diagram (Figure 2), but it does not include any pseudocode or algorithm blocks.
Open Source Code No The code with the network architecture and training can not be currently released under an open-source license that allows others to use, modify, and distribute it due to specifics of the industrial collaboration in the scope of which the work has been done.
Open Datasets Yes Extended versions of two existing CAD datasets ABCParts Li et al. [2019] and CC3D Cherenkova et al. [2020]. The new versions, called ABC-VEF and CC3D-VEF, include the proposed LAR-based point-to-BRep adjacency representation on the point clouds and will be made publicly available to enable further research;
Dataset Splits Yes ABCParts-VEF Dataset: Spels Net is trained and evaluated on the ABCParts dataset Li et al. [2019] using the same train (22k), test (3.5k) and validation (3.5k) splits.
Hardware Specification Yes The training takes approximately 10 days on a node with 4 Nvidia A100(40Gb) GPUs.
Software Dependencies No The paper mentions software components like 'Sparse CNN encoder' and 'Adam W solver' and 'HDBScan', but it does not specify version numbers for any of these or other underlying software dependencies (e.g., Python, PyTorch/TensorFlow versions).
Experiment Setup Yes The input point cloud is normalized to unit sphere, randomly rotated and discretized on a voxel grid with a resolution ρ = 0.01. Spels Net is trained with Adam W solver with a cosine annealing learning rate schedule starting at 10 3 and weight decay 10 2 for 250 epochs to convergence. In order to facilitate the learning, we set the number of edges to Ne = 128 and faces to Nf = 128.