Data-Efficient 3D Gaussian Splatting
for Sparse-View Reconstruction

Dong-Won Lee
Kyung Hee University
In progress · Mar 2026 – Present
Homepage
3D Gaussian Splatting reconstruction

Given a fixed budget of input frames, which frames matter most for reconstruction?

Overview

3D Gaussian Splatting (3DGS) delivers high-quality, real-time novel-view synthesis, but its reconstruction quality degrades sharply when only a handful of input views are available. Capturing dense, well-distributed views is expensive and often impractical in the field, so the views we do have should be chosen as carefully as possible.

I approach input-frame selection as a dataset-distillation problem — distilling a large pool of candidate views down to a compact, highly informative subset that preserves reconstruction quality under sparse-view settings. The goal is data-efficient novel-view synthesis: comparable rendering quality from far fewer, better-chosen views.

Approach

  • Frame the input-view budget as dataset distillation: select a compact subset of frames that maximizes downstream reconstruction quality, rather than training on every available view.
  • Build on the gsplat framework for 3DGS optimization, with the Mip-NeRF 360 dataset as the primary benchmark.
  • Evaluate reconstruction with standard novel-view-synthesis metrics (PSNR, SSIM, LPIPS) against uniform and random frame-selection baselines.

Status

This is an ongoing research project. I am currently running reconstruction experiments on Mip-NeRF 360 scenes and comparing selection strategies under varying view budgets. Results and code will be added here as the work matures.