Nexel-SLAM: Decoupling Geometry and Appearance for Dense RGB-D SLAM

High appearance complexity need not imply high geometric complexity, yet splatting-based SLAM often ties appearance detail to geometric discretisation. We introduce NeXeL-SLAM, an RGB-D SLAM system that combines explicit surface geometry with spatially varying neural appearance. Building on the nexel representation, we initialise oriented surfels from depth and augment their base colours with a shared neural texture field. Coarse world-space features share information across surfels, while fine surfel-local features attach texture to the evolving geometry. A common differentiable rasteriser supports tracking from depth and base colours, using analytic $\mathrm{SE}(3)$ pose gradients, and mapping with the full neural appearance. New surfels are introduced according to coverage and depth disagreement, while geometry and appearance are jointly refined over keyframes. On Replica, NeXeL-SLAM achieves $37.96\mathrm{dB}$ training-view PSNR, $0.59\mathrm{cm}$ mean Chamfer distance for the fused reconstruction, and $0.28\mathrm{cm}$ all-frame ATE-RMSE. On TUM RGB-D, it achieves $24.31\mathrm{dB}$ training-view PSNR, although fused-surface accuracy is less consistent. These results show that an explicit surface map can support detailed appearance alongside accurate reconstruction and tracking, while identifying remaining challenges in fusing real-sensor observations.