System Evaluation

Target Localization Experiment

Experimental Setup

To evaluate the depth estimation accuracy of FoveaCam Duo, we designed a controlled target localization experiment. A small Statue of Liberty model is used as the target object and moved along a 3D-printed rail. The "∞" shaped rail contains slots at predefined positions, providing ground-truth locations of the target at each slot relative to the geometric center of the rail.

We place the rail at three distances from the camera system: $Z=5$, $10$, and $15$ m, with the plane of the "∞" pattern oriented to face the camera. At each distance, the target is placed sequentially at every slot, and the foveated cameras are steered to center the target in their fields of view.

As a baseline, we evaluate a conventional wide-angle stereo pair with the same baseline $B$ and the same total field of view as the foveated cameras' steerable field of regard. This ensures a fair comparison: both systems observe the same angular extent but differ in how they distribute angular resolution across that extent.

---

Figure 1

Target localization experiment setup. The target, a Statue of Liberty model, is placed at predefined slots along a 3D-printed ∞-shaped rail. The camera system is mounted on a linear rail that translates the center wide camera along the $x$-axis to the positions of the left and right foveated cameras, creating a virtual wide-angle stereo pair at exactly the same location and baseline for fair comparison.

Depth Estimation via Template Matching

For each stereo pair (foveated or wide-angle), we localize the target using normalized cross-correlation (NCC) template matching. A template of the Statue of Liberty model is manually extracted from one image and matched in the other to obtain corresponding pixel coordinates. Sub-pixel refinement with up to $\frac{1}{256}$-pixel precision is then applied to fine-adjust the observed disparity at the center of the matched tile. Given the rectified parallel stereo geometry established by our calibration pipeline for the foveated pair, and by standard stereo calibration for the wide-angle pair, we compute depth via triangulation from the horizontal disparity $d$:

$$Z=\frac{f\cdot B}{d}$$

where $f$ is the focal length in pixels, $B$ is the stereo baseline, and $d$ is the measured disparity.

Localization Results

Target Localization at

Auto Rotate

---

Figure 2

Estimated target trajectories on the ∞-shaped rail at $Z=5$, $10$, and $15$ m for wide-angle stereo (■) and foveated stereo (●). Beyond $10$ m, wide-angle stereo fails to recover a coherent trajectory, while FoveaCam Duo preserves the characteristic ∞ shape throughout. Each estimated trajectory is rigidly translated to align its centroid with ground truth for visualization.

We are still working on improving the accuracy of FoveaCam Duo. The results shown here are subject to change.

At $Z=10$ m, the wide-angle stereo system shows significantly degraded accuracy as the target occupies fewer pixels and the disparity shrinks, quantizing depth into coarse discrete levels. The foveated stereo system, benefiting from its $\sim 9\times$ higher angular resolution, maintains substantially larger disparity and correspondingly lower depth error. At $Z=15$ m, depth from wide-angle stereo fails entirely — the disparity approaches the noise floor of the matching algorithm and the reconstructed trajectory bears no resemblance to the ground-truth "∞" pattern. In contrast, the foveated stereo system continues to produce coherent depth estimates and successfully recovers the characteristic "∞" shape of the target trajectory.

---

Figure 3

Quantitative target localization error at $Z=5$, $10$, and $15$ m (lower is better). FoveaCam Duo maintains low depth error across all distances, while the wide-angle stereo baseline degrades sharply beyond $10$ m.

We are still working on improving the accuracy of FoveaCam Duo. The results shown here are subject to change.

These results confirm the fundamental advantage of foveated stereo: by concentrating angular resolution on the region of interest via MEMS mirror steering, the system maintains the effective disparity needed for accurate triangulation at long range, without sacrificing the ability to cover a wide field of regard.