Dongho Lee — Portfolio

Multi-scale 3DGS renders one scene at many zoom levels (1× → 128×). Gaussians fitted at full resolution break down when the camera zooms out: colors wash out, thin structures alias, and errors pile up at coarse scales. These two works attack that failure from different sides.

SCIE · Published Mathematics (MDPI) · 2026.07 · H. Park†, Y. Lee†, D. Lee, D. Jun*

Adaptive-Scale Color Offset and Error-Guided Gaussian Reallocation for Multi-Scale 3D Gaussian Splatting

ACO-EGR gives every Gaussian a small per-scale color correction and moves under-used Gaussians to where coarse-scale error stays high — without adding primitives.

+2.09dB

PSNR gain at 128× over MS-GS (Mip-NeRF 360)

+1.96dB

PSNR gain at 4× (Mip-NeRF 360)

Benchmarks improved from 4× onward: Mip-NeRF 360, Tanks & Temples, Deep Blending

±0

Extra Gaussians — reallocation runs under a fixed primitive budget

Problem At coarse scales one pixel averages many Gaussians, so a single fixed color per Gaussian is the wrong color once you zoom out.

Idea (a) ACOAdaptive-scale Color OffsetLearns a per-scale DC color offset Δ^s for each Gaussian — geometry stays untouched.(b) EGRError-guided Gaussian ReallocationMoves low-contribution Gaussians into regions where coarse-scale error stays high.

Result Higher PSNR and SSIM at every scale from 4× upward on all three datasets, with rendering time close to MS-GS.

Overview of ACO-EGR: (a) adaptive-scale color offset adds a per-scale DC color offset to each projected Gaussian; (b) error-guided reallocation moves low-contribution Gaussians toward high-error regions. — Overview of ACO-EGR. Fig. 2 from Park et al., *Mathematics* 2026 (CC BY 4.0).

PSNR at every zoom level

Higher is better · hover for values

MS-GSOurs (ACO-EGR)

Drag to compare — MS-GS vs. Ours

Same view rendered at a lower resolution; the boxed inset is a zoomed crop.

MS-GSOurs

↔

Ground truth

Vanilla 3DGS

3DGS

Crops from Fig. 3 of Park et al., Mathematics 2026 (CC BY 4.0).

M.S. Thesis Dong-A University · 2026 · Advisor: Prof. Dongsan Jun

Multi-Scale 3D Gaussian Splatting with Region-Adaptive Structure Loss and Scale-Wise Weighting (RSIC-GS)

Prior anti-aliasing work fixes the renderer. RSIC-GS instead changes the training objective, so that structure-rich regions and coarse scales get the supervision they were missing — as a plug-in for existing backbones.

+0.51dB

Average PSNR gain over MS-GS across 1×–64× (Mip-NeRF 360)

+0.84dB

Largest single-scale gain, at 16×

Backbones it plugs into: MS-GS and Mip-Splatting

−2.3MB

Model size vs. MS-GS on Mip-NeRF 360 — no memory overhead

Problem A single loss treats every region and scale alike: structure-rich regions are under-supervised, and coarse scales dominate the summed loss while fine scales converge early.

Idea RASLRegion-Adaptive Structure LossLaplacian-domain residual between rendered and ground-truth images, auto-normalized per scale.SAWScale-Adaptive WeightingLearnable per-scale weights that rebalance the multi-scale loss.

Result Gains concentrate where MS-GS is weakest (4×–64×), at near-unchanged rendering time; ablation shows RASL and SAW are complementary (25.62 → 26.13 dB).

RSIC-GS training pipeline: (a) RASL computes Laplacian maps of rendered and ground-truth images and their residual; (b) SAW learns per-scale weights through a softplus parameterization. — Training pipeline — (a) RASL, (b) SAW.

Gain over the MS-GS backbone

Mip-NeRF 360 · PSNR (dB) · hover for values

MS-GSMS-GS + RSIC-GS

Bonsai scene: ground truth and zoomed crops at 1×, 4×, 16× and 64× for MS-GS and MS-GS + Ours. — bonsai (Mip-NeRF 360) — zoomed crops at 1×, 4×, 16×, 64×: MS-GS vs. MS-GS + Ours.

Dongho
Lee

About Me

Research Journey

Why 3DGS — the field moved fast

Research Highlights