GeoTrack-GS: Track-Guided Geometry Refinement for Sparse-View 3D Gaussian Splatting

Abstract

Three-dimensional Gaussian Splatting (3DGS) can degrade severely in sparse-view settings because the optimization is under-constrained. Monocular depth priors provide dense supervision, but their inconsistency across views can distort geometry. We propose GeoTrack-GS, a geometry-first framework. We use monocular depth as a fixed, dense weak regularizer for Gaussian geometry. Sparse feature tracks enforce macro-level reprojection consistency, while a micro-level K-NN PCA regularizer suppresses rank collapse. These constraints refine the Gaussian geometry during optimization without explicitly updating the input depth maps. On the optimized Gaussian geometry, GT-DCA captures view-dependent appearance; its geometry-conditioned query is stop-gradient isolated so that the GT-DCA branch does not introduce an additional appearance-to-geometry gradient path through query construction. A Decoupled Constraint Stabilization strategy modulates the geometric and pseudo-view constraints during training. Across LLFF and DTU with 3–9 input views, and Mip-NeRF 360 with 12 input views, GeoTrack-GS maintains competitive rendering quality. It also improves geometric metrics under the evaluated DTU protocol. This geometric conclusion is restricted to DTU, the benchmark in which ground-truth geometry is quantitatively evaluated.