New FOSSA architecture achieves 55.7% error reduction in zero-shot depth estimation

Researchers have developed a novel Transformer-based architecture called FOSSA for the task of estimating dense metric depth maps from focus stacks, known as Depth from Defocus (DfD). This new model emphasizes zero-shot generalization, aiming to perform well on unseen datasets without specific overfitting. To support this, a new benchmark named ZEDD was created, featuring a significantly larger number of scenes and higher quality data than previous benchmarks. The FOSSA architecture incorporates a stack attention layer with focus distance embedding for efficient cross-stack information exchange, and experiments show it reduces errors by up to 55.7% on the ZEDD benchmark. AI

IMPACT Introduces a novel architecture and benchmark for depth estimation, potentially improving scene understanding in computer vision applications.

RANK_REASON This is a research paper detailing a new model architecture and benchmark for a computer vision task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• FOSSA
• Transformer
• Yiming Zuo
• ZEDD

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →