AI & text-to-CAD
DeepCAD dataset
Definition
The DeepCAD dataset is a public research dataset of 178,238 CAD models stored as sketch-and-extrude construction sequences, released with the ICCV 2021 DeepCAD paper. Much text-to-CAD research trains and benchmarks on it, and its simple geometry limits what those models learn.
Also known as: DeepCAD
Updated
What it contains#
The DeepCAD paper introduced a generative network for CAD construction sequences and, to train it, a new dataset of 178,238 models with their construction sequences. Code and data are on GitHub.
The operations are mostly sketches and extrusions. That makes the dataset useful for learning how simple prismatic parts are built, but it has little coverage of fillets, shells, sweeps, patterns, assemblies, or manufacturing detail. Text2CAD later added text descriptions to roughly 170,000 of these models.
Sources
Related terms
Text2CAD
Text2CAD is a NeurIPS 2024 spotlight paper and dataset that generates parametric sketch-and-extrude CAD sequences from text prompts. It added about 660,000 text annotations, from beginner to expert level, to about 170,000 DeepCAD models.
Neural CAD
Neural CAD refers to approaches where neural networks generate CAD modeling operations or editable CAD geometry, rather than meshes or images. Autodesk also uses the name for its work toward turning natural-language prompts into editable design geometry in Fusion.
Feature tree
A feature tree is the ordered list of operations, such as sketches, extrudes, cuts, fillets, and patterns, that builds a parametric CAD model. Editing an early feature replays every feature after it, which makes controlled changes possible and also makes the model depend on each earlier step.