Is Fusion 360 Reconstruction Dataset Really Not Enough for Training When Compared With DeepCAD Dataset?

Area of Science:

• Computer-Aided Design (CAD)
• Machine Learning
• Data Science

Background:

• DeepCAD dataset (178k samples) is widely adopted for CAD sequence learning.
• The Fusion 360 Reconstruction dataset (less than 10k samples) is considered insufficient for generalization.
• Existing research suggests larger datasets are crucial for effective CAD sequence learning models.

Purpose of the Study:

• To investigate and compare the efficacy of the DeepCAD and Fusion 360 Reconstruction datasets in CAD sequence learning.
• To determine if the significant size difference impacts model generalization capabilities.
• To identify future directions for improving CAD datasets.

Main Methods:

• Comparative analysis of two distinct CAD datasets: DeepCAD and Fusion 360 Reconstruction.
• Development and application of reasonable experimental designs.
• Implementation of a data augmentation technique to assess dataset equivalence.

Main Results:

• Both Fusion 360 Reconstruction and DeepCAD datasets exhibit comparable performance in CAD sequence learning for simple sketch and extrusion commands.
• The substantial difference in sample size does not translate to a significant difference in learning capability for basic operations.
• The datasets are found to be essentially indistinguishable in their current application.

Conclusions:

• Dataset size alone is not the sole determinant of effectiveness in CAD sequence learning for basic commands.
• The current datasets may limit advancements due to their simplicity.
• The development of more complex and advanced CAD datasets is essential for future progress in the field.