VideoCAD contains over 41,000 annotated video recordings of 3D CAD UI interactions, developed by Brandon Man at the DeCoDE Lab and updated in 2026. The data is generated via an automated framework that extracts high-fidelity UI action sequences from human-made CAD designs to support long-horizon learning.
Use Cases
- Predicting discrete CAD UI actions from raw video frames
- Modeling long-horizon dependencies in complex 3D engineering workflows
- Benchmarking multimodal models on high-fidelity UI interaction sequences
Strengths
- 41,000+ annotated video recordings
- 20x longer time horizon than existing CAD interaction datasets
- High-fidelity UI action labels derived from human-made designs
Limitations
- Synthetic generation may introduce a domain gap compared to organic screen recordings of human users
- The specific CAD software environment used for generation is not explicitly named in the metadata
Provenance
- Source
- DeCoDE Lab
- Collection Method
- synthetic
- Freshness
- Last updated March 2026.