Same Measurements, Different Velocity Risk: Complementary Rolling-Shutter Readouts under Clock Uncertainty

Lunshuo Tian, Dongheng Liu, Mingwei MaarXiv 2026 · arXiv

When you’re trying to reconstruct the 3D world from a moving camera, timing is everything. For rolling-shutter (RS) cameras—the kind in almost every smartphone—this is doubly true because they capture images row-by-row. If you don't know exactly when each row was exposed (clock uncertainty), your velocity estimates can go haywire.

A new paper from researchers at UNSW Sydney, Sorbonne University, and South China University of Technology reveals a surprising fix: simply flip the readout direction of one camera. This "complementary" design can fundamentally change the statistical difficulty of the problem, making velocity estimation far more robust without adding a single extra pixel to your data budget.

TL;DR

In rolling-shutter camera rigs, "repeated" readouts (where all cameras scan in the same direction) suffer from a fundamental ambiguity between motion and timing that creates a high "velocity risk." By using "complementary" readouts (reversing one camera's scan direction), you break this symmetry, drastically improving the accuracy and stability of velocity estimation under uncertain exposure timing.

What Problem Does This Paper Address?

Rolling-shutter cameras introduce distortions because the camera moves while the image is being scanned. To "un-distort" the image and estimate the camera's velocity, you need precise timestamps for every row. However, in multi-camera rigs, clocks are often un-synchronized or have "jitter" (uncertainty).

The core challenge is the velocity-timing ambiguity: a specific change in camera velocity can look identical to a specific shift in exposure timing. To an optimizer, these two very different physical realities produce the same image data. The authors wanted to know if the physical layout of the cameras—specifically their readout directions—could be designed to eliminate this ambiguity.

Methods

The authors compared two acquisition designs on a four-camera rig:

  1. Repeated Readouts (+, +): All auxiliary cameras scan in the same direction (e.g., top-to-bottom).
  2. Complementary Readouts (+, -): One auxiliary camera scans in the opposite direction (bottom-to-top).
Figure 1. We compare two readouts on the same four-camera rig. Anchors A0/A1 have known relative calibration; auxiliaries U1/U2 have independently unknown poses and clocks. We show the reference centres in normalized body coordinates, with parallel optical axes along +z. We reverse only the nominal roll of U2; arrows show increasing exposure time. This is a schematic of the theoretical reference.

They used a rigorous statistical framework to analyze "worst-case velocity risk." They proved that for repeated readouts, there exists a "compensation path" where timing errors and velocity changes perfectly cancel each other out to first order.

To prove the stability of the complementary design, they developed a nonlinear velocity inverse, showing that even with unknown poses, clocks, and scene structure, the complementary geometry prevents these errors from hiding in the data.

Datasets

The study relies on controlled synthetic environments to isolate the effects of geometry from other noise sources:

  • 32 Synthetic Scenes: A mix of 16 standard and 16 "challenging" scenes (with varying depths, contrast, and motion).
  • Raster Studies: High-fidelity renders of ArUco markers were used to test the full image-processing pipeline, including corner detection and quantization noise.
  • Certified Domains: They defined specific "physical classes" (sets of possible motions and geometries) to provide mathematical guarantees on the error rates.

Key Insights and Ideas

  • The "Clock-Jet" vs. "Velocity-Jet": In repeated readouts, the way an image changes due to a clock shift (the clock-jet) is collinear with the way it changes due to a velocity shift (the velocity-jet). They cancel out. In complementary readouts, these "jets" point in different directions, making them distinguishable.
Figure 2a. Schematic showing how image changes from timing ($d_\tau$) and physics ($d_{phys}$) cancel in repeated layouts but not in complementary ones.
  • Risk Scaling: The paper establishes a massive theoretical gap. As image noise ($\sigma$) goes to zero, the risk for repeated readouts stays high (order $\sigma$), while for complementary readouts, it drops much faster (order $\sigma^2$).
  • Mounting Geometry is a "Lift": Interestingly, the authors found that certain mounting choices (like a common depth offset between cameras) can also help "lift" the ambiguity, but the complementary readout is a more fundamental and robust solution.

Key Results

  • RMSE Reduction: In raster experiments with 2-DN noise and 1-ms exposure, the complementary layout achieved a 39.1% reduction in velocity RMSE compared to the best-performing repeated layout.
  • Convergence Stability: Complementary designs were significantly more "regular." In one study, 100% of complementary fits terminated successfully, whereas less than 50% of repeated-layout fits reached a stable solution.
  • Sensitivity to Noise: As shown in the "Finite image gap" plot, the complementary design maintains a measurable "gap" in the image data even when motion is small, whereas the repeated design's gap collapses, leading to high uncertainty.
Figure 2b. Plot showing the "image gap" remains high for complementary layouts, providing a signal for the optimizer even at low motion amplitudes.

Limitations and Future Work

  • Local vs. Global: The mathematical proofs are "local," meaning they apply to small neighborhoods of motion and geometry. While the experiments suggest broader applicability, a global minimax proof remains elusive.
  • Synthetic to Real: While the raster studies are high-fidelity, the paper does not test on real-world hardware with physical clock jitter. Real sensors may have additional non-idealities (like rolling-shutter wobble or non-linear readout speeds) not captured here.
  • Rig Specificity: The analysis is tailored to a four-camera rig. While the principle of "readout diversity" is general, the exact benefits for other configurations (e.g., a single-camera smartphone moving in a specific way) require further study.

The takeaway for engineers is clear: if you're building a multi-camera system for high-speed robotics or SfM, don't just align your sensors—consider flipping one over. It might just be the cheapest accuracy boost you'll ever get.

Summary updated

Abstract

Camera rigs with identical measurement budgets can have different velocity estimation rates when exposure timing is uncertain. We establish this distinction for rolling-shutter reconstruction with two calibrated moving anchors, unknown scene points, and independently unknown auxiliary poses and clocks. On the same fixed local physical class near a separated lateral rig, repeated readouts have worst-case squared velocity risk at least of order $σρ_{\mathrm{eff}}$, whereas complementary readouts have minimax risk of order $σ^2$, for a fixed admissible clock displacement $ρ_{\mathrm{eff}}>0$ as the noise standard deviation $σ$ tends to zero. A finite physical compensation path gives the lower bound; a nonlinear velocity inverse proves complementary stability. All motion, structure, pose, and clock nuisances remain unknown. We identify mounting configurations in which both designs become regular and quantify the risk floor induced by metric uncertainty. We quantify full-parameter domains with independent 6-ms clocks and test in-domain reconstruction under paired noise and a depth-offset intervention. Raster studies examine image formation and calibration sensitivity.
BibTeX
@misc{tian2026same,
  author = {Lunshuo Tian and Dongheng Liu and Mingwei Ma},
  title = {{Same Measurements, Different Velocity Risk: Complementary Rolling-Shutter Readouts under Clock Uncertainty}},
  year = {2026},
  eprint = {2611.10774},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10774},
}
Figure