Macro3DGS: Sharp 3D Gaussian Splatting of Macro-Scale Subjects from Sparse Focal Slices

抄録

Macro-scale objects, such as insect specimens, contain rich surface details that invite close examination from arbitrary viewpoints. Macro photography, however, suffers from an extremely shallow depth of field, leaving only a small portion of each image in focus. Consequently, 3D reconstruction methods such as 3D Gaussian Splatting (3DGS) average focused and defocused observations into blur. Existing pipelines therefore merge tens of focal slices per viewpoint into all-in-focus images before reconstruction, requiring over a thousand photographs for a single specimen. Our key insight is that focus stacking is fundamentally a per-pixel, sharpness-weighted selection over a focus stack that can be performed during 3DGS optimization rather than as image preprocessing. Building on this insight, we propose Macro3DGS, which weights the photometric loss by a softmax-normalized focus measure across each focal slices, warps the slices to cancel focus breathing, and prunes near-camera floaters. Our experiments reveal that Macro3DGS requires only sparse focal slices: using three slices per viewpoint, and it renders sharper novel views than anti-aliased and defocus-aware baselines.