Revisiting Pose Sensitivity in Splat-based Computed Tomography under Sparse-view Reconstruction

Splat-based tomography has been attractive because it converges quickly and can produce high-quality sparse-view reconstructions by modeling volume as continuous 3D Gaussians. But in real CT setups with limited, uneven camera angles, the technique can fall apart in a very specific way: pose inaccuracies in the acquisition geometry trigger visible streak and strip artifacts.

The key shift here is the diagnosis. Instead of treating sparse views as the main culprit, the work argues that geometric misalignment is the bigger problem for this formulation. That matters for anyone building differentiable reconstruction systems, because it changes where the engineering effort should go: not just more views or heavier regularization, but tighter pose handling and joint optimization.

To address that, the method revisits pose sensitivity in the splatting pipeline and introduces a stable gradient-based framework that refines geometric parameters during reconstruction. The authors also analyze how pose perturbations flow through the differentiable projection operator, which helps explain why splat-based CT is unusually vulnerable to misalignment compared with more conventional approaches.

For developers, the practical appeal is that the fix is described as lightweight and easy to drop into existing pipelines. If it holds up beyond the lab, it could make splat-based reconstruction more viable in real-world sparse-view scenarios where perfect calibration is unrealistic and geometry drift is hard to avoid.