- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes Masked Topology Modeling (MTM), a self-supervised learning approach for learning representations of parametric CAD models in B-rep format. The method represents a B-rep as a face-adjacency graph, masks some of the connections between faces, and trains an encoder to predict properties of the masked edges, including their convexity and curve type. Since these properties can be obtained directly from the B-rep, the model can be pretrained without manually labeled data. The authors combine MTM with contrastive learning and evaluate the learned representation on several downstream CAD classification and segmentation tasks. The experiments show that MTM is particularly effective when only a small amount of labeled data is available, while its advantage becomes smaller when sufficient labeled training data is used. Overall, the paper demonstrates that the intrinsic topology of B-rep models can be used as a self-supervised signal to improve the label efficiency of CAD representation learning.
Critical strengths
The paper addresses an interesting and important problem of learning useful representations for CAD models when labeled data is limited. The observation that masking-based self-supervised learning, which has been successful in language and image domains, can also be effectively applied to the topology of B-rep models is interesting. In particular, predicting properties of masked edges provides a natural way of using the information already available in CAD models as supervision. Another strength is the extensiveness of the experiments. The authors evaluate the approach on multiple datasets, downstream tasks, architectures, and different amounts of labeled data, and provide several ablation studies to understand the contribution of different components of the method.
Critical weaknesses
My main concern is the practicality of the proposed method. The largest improvements are observed in extremely low-data settings, such as using only 0.1% of the labeled data. While these experiments demonstrate the label efficiency of the learned representation, it is unclear how representative such settings are of practical CAD applications, where labeled B-rep datasets are available and additional synthetic data can also be generated. When more labeled data is available, the advantage of the proposed method becomes considerably smaller and in some experiments it does not outperform existing approaches. This raises the question of whether the additional pretraining stage provides a significant practical benefit outside very limited-data scenarios.
Questions for authors
The largest improvements seem to happen in extremely low-data settings, such as 0.1% labeled data. What practical CAD applications do the authors envision where labeled data is this scarce?
When more labeled data is available, the improvement from MTM becomes relatively small and sometimes disappears. How do the authors justify the additional pretraining cost in these settings?
Given that synthetic B-rep data can be procedurally generated, why is self-supervised pretraining preferable to simply generating more labeled or automatically annotated training data?
Masking-based representation learning has been extensively explored in language and vision. Beyond adapting masking to B-rep topology, what do the authors consider the main conceptual novelty of MTM?
The method predicts convexity and curve type for masked edges. Why were these two properties selected?