Photogrammetric Record Инженерия (прочее)

Mitigating Weak‐Parallax Degeneracy in Aerial Bundle Adjustment via Parallax Angle Parameterization

Wenyi Lu, Zhengkang Zuo, Xiang Zhang, Yanbiao Sun, Yajing Wu, Chia Ming Toh, Lizhe Wang, Mei Xue, Lei Yan, Yunfan Dai
Wiley · 2026-10-01 · DOI 10.1111/phor.70072

Аннотация

ABSTRACT Weak‐parallax aerial image blocks, common in forward‐motion oblique sequences and long‐range nadir datasets, often degrade the numerical conditioning and convergence reliability of bundle adjustment (BA) in aerial photogrammetry. We therefore benchmark parallax angle parameterization (PAP) against Euclidean XYZ and inverse depth parameterizations in a strictly controlled BA backend, quantifying its impact on conditioning and convergence behavior. Using the same camera calibration model, right‐multiplicative SO(3) pose updates, and analytical Jacobians across all settings, we conduct a controlled 3 × 2 × 2 full‐factorial study on 12 real aerial datasets, including three oblique‐view datasets and nine nadir‐view datasets, in which we vary the parameterization, compare the Gauss–Newton (GN) and Levenberg–Marquardt (LM) nonlinear solvers, and assess the impact of robustification by turning the robust loss on or off. We report the mean squared error (MSE) of reprojection, iteration counts, and total BA runtime, and subsequently summarize factor‐wise trends using within‐dataset paired deltas aggregated across datasets. Results show that parallax angle bundle adjustment (PA‐BA) exhibits a smoother and faster convergence trajectory with steeper MSE decay and reduced oscillations, leading to the highest convergence success rates of 100% in oblique scenarios and 95% in nadir scenarios. In the full‐factorial comparison, paired‐delta aggregation shows that the most consistent gains emerge under LM and are further stabilized by Huber robustification, primarily reducing iteration demand and total BA runtime. When optimization converges to the same global optimal region, all parameterizations reach essentially comparable minima, with final MSE differences typically within ±10%.

Читать у издателя →