11-20

Yiming Zuo FPO

Yiming Zuo will present his FPO "Monocular Depth Estimation with Auxiliary Measurements" on Friday, November 20, 2026 at 8:30 AM in Friend 202.

 

The members of Yiming’s committee are as follows:

Examiners: Jia Deng (Adviser), Adam Finkelstein, Felix Heide

Readers: Szymon Rusinkiewicz, Zhuang Liu

 

A copy of his thesis is available upon request. Please email gradinfo@cs.princeton.edu if you would like a copy of the thesis.

 

Everyone is invited to attend the talk.

 

Abstract follows below:

 

Monocular depth estimation predicts dense scene geometry from a single image, which supports applications in robotics, 3Dreconstruction, and XR. Despite its importance, estimating dense metric depth from a single image is challenging due to inherent ambiguity: a larger object farther from the camera can produce the same image as a smaller object closer to the camera. Thankfully, auxiliary measurements are often available in real-world image capture systems, either externally (e.g., sparse depth from LiDAR) or internally (e.g., changes in defocus blur as a camera adjusts its focus). This thesis studies how to combine these measurements with learned visual priors to achieve accurate and robust depth estimation, with an emphasis on generalization to unseen datasets and measurement conditions.

 

We first introduce OGNI-DC (ECCV 2024), a depth completion framework based on Optimization-Guided Neural Iterations. A recurrent network refines a depth gradient field, and a Differentiable Depth Integrator converts these gradients into dense depth under constraints from sparse observations. This design combines strong in-domain accuracy with robustness across datasets and sparsity levels. We then introduce OMNI-DC (ICCV 2025), which addresses error accumulation in depth integration when observations are extremely sparse. We introduce a Multi-resolution Depth Integrator to model long-range depth relationships, together with a Laplacian loss, scale normalization, and diverse synthetic depth patterns, allowing a single model to perform zero-shot depth completion across datasets and sensors.

 

Finally, in Zero-Shot Depth from Defocus (ECCV 2026), we study the use of a focus stack and the corresponding focus distances to predict dense metric depth. We propose FOSSA, a Focus Stack Attention Transformer that extracts defocus cues through stack attention and fuses the features early for efficient prediction. Training on synthetically generated focus stacks enables generalization to real images. We also introduce ZEDD, a real-world benchmark with diverse scenes, multiple apertures, and high-quality depth ground truth.

 

Together, these contributions show how model designs tailored to auxiliary measurements, combined with diverse training data, improve the accuracy and generalization of monocular depth estimation.