Masked BRep Autoencoder via Hierarchical Graph Transformer
TLDR
Self-supervised masked BRep autoencoder with hierarchical graph transformer learns CAD representations for downstream tasks with limited labeled data.
Reasoning
The paper presents a novel self-supervised framework combining masked autoencoding and hierarchical graph transformers for BRep models, achieving strong performance on classification, segmentation, and feature recognition. Strengths include the innovative architecture and effective use of unlabeled data; weaknesses are the lack of explicit real-world dataset details and limited scope to representation learning rather than generation.
Read-first score
Read-first score 52.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.
Field roles
Rank sensitivity
Stability: volatile; rank range: 27.
Keyword Scores
Deep Analysis
Innovations
- Masked graph autoencoder that reconstructs randomly masked geometries and attributes of BReps for self-supervised representation learning
- Hierarchical graph Transformer architecture with cross-scale mutual attention block for long-range geometric dependencies and GNN block for local topological information fusion
Methodology
A masked graph autoencoder is pretrained on a large-scale unlabeled BRep dataset by randomly masking and reconstructing geometries and attributes. The encoder uses a hierarchical graph Transformer that combines a cross-scale mutual attention block to capture long-range dependencies and a GNN block for local topology. After pretraining, the decoder is replaced with a task-specific network fine-tuned on small labeled data for downstream tasks.
Key Results
The model achieves high performance on part classification, modeling segmentation, and machining feature recognition, significantly outperforming other methods especially when labeled training data is very limited.