Layer-Based 3D Gaussian Splatting for Sparse-View CT Reconstruction

ICLR 2026 Conference Submission845 Authors

Keywords: CT Scan, Gaussian Splatting, 3D reconstruction, Computer Vision

Abstract: We introduce a dynamic framework for 3D sparse-view Gaussian Splatting that learns scene representations through layerwise, iterative refinement of the Gaussian primitives. Conventional methods typically rely on dense, one-time initialization, where the placement of Gaussians is guided by 2D projection supervision and density control. However, such strategies can lead to misalignment with the true 3D structure, particularly in regions with insufficient projection information due to sparse-view acquisition. In contrast, we adopt a coarse-to-fine approach beginning with a base representation and progressively expanding it by adding new layers of smaller Gaussians to accommodate finer-grained details. At each such iteration, the placement of new primitives is guided by a 3D error map, obtained by the back projection of 2D projections' residuals. This process acts as adaptive importance sampling in 3D space, directing model capacity to regions with high error. We evaluate our approach on sparse-view computed tomography reconstruction tasks, demonstrating improved performance over existing methods.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 845