3D Gaussian Splatting (3DGS) can reconstruct photorealistic scenes from posed images, but it assumes a wide field of view and a deep depth of field. Macro lenses shrink both to millimeters, so a small viewpoint error can push the sample out of frame and only a thin focal plane is sharp. Existing automated macro-capture rigs handle shallow depth of field with focus stacking at each viewpoint, yet they typically rely on classical photogrammetry and offline, human-in-the-loop processing.

MicroSplat instead uses a low-cost consumer microscope on a manipulator arm. After a single up-front object-localization step, the system runs unattended: it performs calibration, plans collision-free viewpoints, captures and fuses a focal stack at each viewpoint, segments the object across all views, and optimizes a 3DGS model initialized from the arm's own pose estimates.

The authors compare MicroSplat against blur-aware 3DGS methods that infer sharp scenes from defocused images. They report that measuring sharpness with a controlled focal sweep outperforms inference. The paper also releases real and synthetic datasets, including robot and camera parameters, along with the fitted 3DGS reconstructions.