Adaptive Tiling for Least-Squares Phase Unwrapping: Runtime and Accuracy

The study "Adaptive Tiling for Least-Squares Phase Unwrapping: Runtime and Accuracy" (arXiv:2609.28541v1) investigates whether adaptive tiling (quadtree, kd-tree) improves the speed and accuracy of phase unwrapping compared to a regular grid.

Key Findings: Slower Performance: Optimized adaptive partitions are slower than regular grids. Quadtree reconstruction took 24% longer at minimum tile size $s=8$ and 53% longer at $s=16$ in median paired time. Reason for Overhead: The computational savings from fewer tile boundaries are outweighed by partition construction costs (24–25% of time) and larger local solve work per pixel. Retaining large tiles does not reduce the $O(n_t \log n_t)$ transform work per tile. Accuracy Trade-offs: Adaptive methods generally showed lower accuracy (higher RMSE) and did not consistently improve reference agreement. More refinement did not guarantee better results. Conclusion: The regular grid remains the faster baseline for single-threaded, encoded-input reconstruction. Adaptive tiling is only advantageous if it can provide a faster approximate result at a specified error tolerance, which was not demonstrated in this configuration.

Technical Details: Methods Compared: Regular grid, quadtree, and kd-tree partitions. Criteria Evaluated: Nine subdivision criteria including residue count, fringe density, and phase variation measures. Dataset: 198 entries from a heterogeneous image dataset. Solver: Least-squares reconstruction using Discrete Cosine Transform (DCT) or FFTW.

Problem context. The paper addresses a practical bottleneck in large-scale least-squares phase unwrapping: how to partition a phase image into local reconstruction problems. Phase unwrapping recovers the missing integer multiples of \(2\pi\), and least-squares formulations can be solved more efficiently by working on tiles rather than the full image. Uniform tiling is attractive because it bounds problem size and exposes parallelism, but it creates many local subproblems and tile boundaries even in regions where the phase field is smooth and requires little refinement.

Contributions and insights. The central contribution is an investigation of adaptive tiling as a way to reduce both computational cost and the number of tile interfaces. Instead of using a fixed tile size, the approach retains large tiles where local structure is simple and subdivides only where greater resolution or refinement is needed. The paper examines the resulting trade-offs between runtime and unwrapping accuracy, with the motivation that fewer tiles and fewer boundaries can reduce overhead, limit seam-related artifacts, and preserve larger coherent regions during the least-squares solve. The work therefore treats tiling not merely as an implementation detail, but as a design variable that directly affects performance and reconstruction quality.

Why it matters. Phase unwrapping is a key step in many imaging and metrology applications, including interferometric SAR, optical coherence tomography, digital holography, and precision surface measurement, where large images must be processed quickly and reliably. If adaptive tiling can improve runtime without degrading accuracy—or improve accuracy by reducing unnecessary seams—it could make least-squares unwrapping more scalable on parallel hardware and better suited to real-time or large-data pipelines.