Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 4·5 min read
CrossProjection: Geometric Grounding Beyond Viewpoint Change in Architectural Drawings
TL;DR
CrossProjection is a new diagnostic tool that evaluates how well vision-language models maintain component identity and geometry in architectural drawings across different views. It reveals that while some models perform well in closed-choice scenarios, they struggle with free geometric localization.
✦ Why It Matters
Evaluate your vision-language models' performance on architectural drawings to ensure reliable geometric grounding.
Key Takeaways
- Test vision-language models on architectural drawings for component identity preservation.
- Be cautious of relying solely on categorical correctness for geometric reliability.
- Utilize CrossProjection to assess model performance across different architectural views.
- Establish audit trails for evaluating geometric grounding in CAD/BIM applications.
How It Works
CrossProjection evaluates models by analyzing their ability to match, register, and ground geometry across different architectural views. It uses categorical judgments and localization tasks to assess performance, focusing on how well models can identify and locate points, lines, and regions in drawings.