ABC: A Big CAD Model Dataset For Geometric Deep Learning (1812.06216v2)

Abstract: We introduce ABC-Dataset, a collection of one million Computer-Aided Design (CAD) models for research of geometric deep learning methods and applications. Each model is a collection of explicitly parametrized curves and surfaces, providing ground truth for differential quantities, patch segmentation, geometric feature detection, and shape reconstruction. Sampling the parametric descriptions of surfaces and curves allows generating data in different formats and resolutions, enabling fair comparisons for a wide range of geometric learning algorithms. As a use case for our dataset, we perform a large-scale benchmark for estimation of surface normals, comparing existing data driven methods and evaluating their performance against both the ground truth and traditional normal estimation methods.

• The paper introduces a million-model CAD dataset that enables comprehensive benchmarking of geometric deep learning methods.
• The paper details a large-scale benchmark for surface normal estimation, highlighting performance gaps between deep learning and traditional algorithms.
• The paper provides an open-source geometry processing pipeline that supports resampling and precise ground truth comparisons for CAD models.

Analysis of "ABC: A Big CAD Model Dataset For Geometric Deep Learning"

This paper introduces the ABC-Dataset, a robust collection of over one million computer-aided design (CAD) models specifically designed to facilitate research in geometric deep learning. The dataset provides a myriad of explicitly parametrized curves and surfaces, delivering essential ground truth data for tasks such as differential quantities estimation, patch segmentation, geometric feature detection, and shape reconstruction.

Dataset Structure and Benefits

The ABC-Dataset principally espouses a boundary representation (B-Rep) format—one that offers explicit geometric descriptions of solids. This format aptly caters to the generation of numerous data formats and resolutions, thereby enabling equitable comparisons across various geometric learning algorithms. Its representation allows models to be resampled at arbitrary resolutions without error, a significant advantage over existing datasets predominantly comprised of point clouds or meshes.

One notable application of the dataset was a large-scale benchmark for surface normals estimation. This paper highlighted the performance discrepancies between existing deep learning methods and traditional algorithms when compared against ground truth data, providing key insights into their respective efficacy.

Contributions and Implementation

The authors contribute significantly through three main vectors:

1. Dataset Creation: They assemble a distinguished dataset for geometric deep learning, constituting over a million geometric models. This collection is rich with parametric surfaces and features, along with robust ground truth information relevant to patches, sharp features, and differential properties.
2. Benchmark Development: A thorough evaluation is conducted through a benchmark specifically focused on surface normal estimation. This benchmark underscores both local and global methodologies, facilitating a comprehensive analysis of deep learning algorithms while providing a baseline for future developments.
3. Processing Utilities: The provision of an open-source geometry processing pipeline is invaluable, allowing researchers to process CAD models in formats conducive to deep learning tasks. Continual updates to the dataset and benchmarks will ensure its relevance and applicability in ongoing research.

Experimental Insights

The experiments revealed several noteworthy points:

• Data versus Traditional Methods: In instances involving point clouds, data-driven methods exhibited significant promise over traditional geometric approaches. However, when connectivity information was leveraged, even the most basic analytic methods outperformed their deep learning counterparts. This suggests that the current graph neural network architectures are struggling to fully exploit the available connectivity information effectively.
• Expansive Evaluation: The benchmark introduced employs a scale not seen in previous studies, providing an unprecedented evaluation of both data-driven and analytic methods across various resolutions and dataset sizes. This large-scale evaluation is crucial as it reflects how methods perform on complex and realistic data that mimic real-world scenarios.
• Importance of Regular Sampling: The structured parametrization supports consistent sampling and accurate ground truth comparison, emphasizing the importance of high-quality data in driving the outcomes of geometric deep learning models.

Implications and Future Directions

The ABC-Dataset stands as a critical advance for geometric deep learning, providing a substantial repository that supports the development and benchmarking of innovative methods. Future work could explore advancing deep learning architectures to more effectively utilize surface connectivity information or enhance robustness against high variability in model geometries. Additionally, expansion towards diverse categories of geometric models could further augment the dataset's utility across various domains.

In conclusion, the ABC-Dataset represents a significant resource, enabling researchers to push the boundaries of what is conceivable in geometric deep learning, paving the way for a deeper understanding and more advanced applications in AI-driven geometry processing.