FoveaCam Duo: Stereo Foveated Camera System for High-quality Long-range Depth Sensing ICCP 2026

Yuxuan Zhang, Jacob Carter, Hannah Kirkland, Noah Ralph, Michael Tomadakis, Sanjeev Koppal

Introduction

We present FoveaCam Duo, a novel stereo camera system equipped with dual telephoto sensors capable of rapidly foveating among different targets. At the demo, we will showcase the working camera system with a custom-developed GUI application for real-time high-resolution stereo tracking and streaming, as well as the full calibration setup including hardware, interfaces, and the calibration process. See Demo Experience for details.

---

Figure 1

Annotated image of the FoveaCam Duo system.

FoveaCam Duo is a biologically inspired foveated stereo camera system. It combines a central wide-angle camera with two MEMS-mirror-steered telephoto cameras, decoupling coverage from resolution: the wide camera covers a large area, and the foveated pair "zooms in" with $\sim 9\times$ the focal length to deliver fine-grained depth patches at any region of interest. The MEMS mirrors redirect gaze in sub-millisecond intervals, enabling real-time high-resolution passive stereo depth across a wide field of regard.

Motivation

---

Figure 2

Comparison between (a) a traditional wide-angle stereo pair with parallel optical axes, (b) vergence stereo with telephoto lenses, and (c) foveated vergence stereo that achieves both high resolution and a wide field of view.

Conventional stereo cameras impose a fundamental trade-off between field of view and depth resolution: wide-angle lenses provide global context but dilute angular resolution, while telephoto optics concentrate detail at the expense of coverage. This trade-off becomes particularly limiting when accurate dense metric depth is required across both near and far ranges. FoveaCam Duo overcomes this limitation by optically separating the two concerns.

Because the steered optics introduce image distortion that varies continuously with mirror angle, traditional fixed-geometry stereo calibration does not apply. We address this with a three-stage calibration pipeline that maps MEMS voltages $(v_x,v_y)$ to pointing angles $({\alpha }_x,{\alpha }_y)$ and rectifies foveated images into a virtual parallel-stereo pinhole model via angle-dependent homography regression.

Contents

🔭

System Design →Hardware overview, optical performance, and depth resolution analysis.

📐

Camera Calibration →Intrinsic, extrinsic, and homography rectification pipeline for angle-varying optics.

📊

Evaluation →Target localization experiment and quantitative depth accuracy results.

🎮

Demo Experience →What attendees will see and interact with at the live demo.