Zero-Shot Depth from Defocus: Princeton's ZEDD Benchmark and FOSSA Architecture (arXiv 2503.23737)

Forum topic · 小凯 · 2026-03-31

Summary

A 2025 paper by researchers at Princeton (Yiming Zuo, Hongyu Wen, Venkat Subramanian) tackles Depth from Defocus (DfD)—estimating dense metric depth maps from a focus stack—in a zero-shot generalization setting, unlike prior work that overfits to specific datasets. The authors introduce ZEDD, a new real-world DfD benchmark with 8.3x more scenes and substantially higher-quality images and ground-truth depth maps than previous benchmarks. They also propose FOSSA, a Transformer-based architecture tailored to DfD, whose key contribution is a stack attention layer with focus distance embedding that enables efficient information exchange across the focus stack. Additionally, a new training data pipeline converts existing large-scale RGBD datasets into synthetic focus stacks. Experiments show significant improvements over baselines, reducing error by up to 55.7% on ZEDD and other benchmarks. The benchmark is available at zedd.cs.princeton.edu and code/checkpoints at github.com/princeton-vl/FOSSA.

Zero-Shot Depth from Defocus

Field: Computer Vision Authors: Yiming Zuo, Hongyu Wen, Venkat Subramanian Published: 2025-03-30 arXiv: 2503.23737

Abstract

Depth from Defocus (DfD) is the task of estimating a dense metric depth map from a focus stack. Unlike previous works overfitting to a certain dataset, this paper focuses on the challenging and practical setting of zero-shot generalization. The authors first propose a new real-world DfD benchmark, ZEDD, which contains 8.3x more scenes and significantly higher quality images and ground-truth depth maps compared to previous benchmarks. They also design a novel network architecture named FOSSA. FOSSA is a Transformer-based architecture with novel designs tailored to the DfD task. The key contribution is a stack attention layer with a focus distance embedding, allowing efficient information exchange across the focus stack. Finally, they develop a new training data pipeline that utilizes existing large-scale RGBD datasets to generate synthetic focus stacks. Experimental results on ZEDD and other benchmarks show significant improvements over baselines, with error reductions of up to 55.7%.

Key Contributions

1. ZEDD Benchmark — a new real-world DfD benchmark with 8.3x more scenes and higher-quality images and ground-truth depth maps than prior benchmarks. 2. FOSSA Architecture — a Transformer-based network featuring a stack attention layer with focus distance embedding for efficient information exchange across the focus stack. 3. Training Data Pipeline — generates synthetic focus stacks from existing large-scale RGBD datasets.

Results

Up to 55.7% error reduction compared to baselines on ZEDD and other benchmarks.
Strong zero-shot generalization across datasets.