Sterkfontein D3 — GS Representation & Loss Study

01 · Protocol

What is being compared

This is a representation/loss pilot, not yet a benchmark claim. The single HDR+ target is held out from training, and all reported still-image metrics are computed on that one view.

89noisy camera-linear sRGB training views

1held-out HDR+ test target

100kfixed maximum Gaussian budget

30koptimization steps per method

Fairness controls. Official reference COLMAP geometry; identical train/test split; identical initialization budget; no tone-mapping mismatch between prediction and target. The trajectory videos use one shared interpolated COLMAP path at 1020×765, 158 frames, and 24 fps.

02 · Inputs

Training and test examples

The training frames contain view-dependent spotlight motion, shadows, clipped highlights, and sensor noise. The test image is the released HDR+ reference at a held-out pose.

Training frames

Training frame 1 — **Train #001**Noisy camera-linear sRGB

Training frame 20 — **Train #020**Different pose and illumination

Training frame 40 — **Train #040**Moving spotlight / shadows

Training frame 60 — **Train #060**View-dependent exposure pattern

Held-out test target

03 · Full view

Same held-out camera, separated images

Unlike the raw trainer output, each card contains only the reconstruction. GT is shown once as its own card, making exposure, boundary, and texture differences easier to see.

Ground truth — **HDR+ target**Held-out reference

NTF plus 2DGS — **NTF + 2DGS**Best PSNR/SSIM among completed methods

Scalar SVF plus 2DGS — **Scalar SVF + 2DGS**Threshold-free variance stabilization

RGB SVF plus 3DGS — **RGB-SVF + 3DGS**Channel-wise calibrated loss

RGB SVF plus 2DGS — **RGB-SVF + 2DGS**Channel-wise calibrated loss

Textured Gaussians — **Standalone Texture-GS**30k final render

04 · Pixel inspection

Three fixed crops

Every crop uses identical source coordinates. Use these panels for rock microtexture, ceiling boundaries, and dark-region stability; the full images above remain available for context.

A · Central illuminated microtexture

High-frequency rock texture under strong spatially varying illumination.

B · Ceiling and occlusion boundary

Fine geometry, dark crevices, and the edge of the spotlight.

C · Dark right-hand structure

Low-signal texture where noise-aware losses should matter most.

05 · Quantitative result

Held-out HDR+ target at 30k

PSNR and SSIM favor pixel alignment; LPIPS is more sensitive to perceptual texture. The Texture-GS row is filled from its final checkpoint.

Method	Representation	Loss	PSNR ↑	SSIM ↑	LPIPS ↓
Vanilla 3DGS	3D Gaussian	RGB	20.339	0.6339	0.5296
Vanilla 2DGS	2D Gaussian	RGB	20.312	0.6339	0.5156
NTF + 2DGS	2D Gaussian	NTF	20.402	0.6364	0.5069
Scalar SVF + 2DGS	2D Gaussian	Scalar SVF	20.257	0.6325	0.5082
RGB-SVF + 3DGS	3D Gaussian	RGB-SVF	20.235	0.6308	0.5196
RGB-SVF + 2DGS	2D Gaussian	RGB-SVF	20.254	0.6324	0.5104
Standalone Texture-GS	Textured Gaussian	RGB	20.208	0.6096	0.4056

Current signal. NTF provides the best pixel-aligned result among the standard GS variants. Texture-GS reaches LPIPS 0.4056 at 30k versus 0.5069 for NTF, suggesting a strong perceptual-texture benefit that PSNR does not capture; its lower PSNR/SSIM also warns that the texture is not perfectly pixel-aligned.

06 · Temporal comparison

One synchronized camera trajectory

These are novel-path diagnostic renders, so there is no GT video. Compare flicker, floaters, geometry tearing, exposure continuity, and whether fine rock texture remains stable under motion.

0:00 / --

Vanilla 3DGSShared 158-frame path

Vanilla 2DGSShared 158-frame path

NTF + 2DGSNoise-aware loss

Scalar SVF + 2DGSThreshold-free transform

RGB-SVF + 3DGSChannel-wise transform

RGB-SVF + 2DGSChannel-wise transform

Standalone Texture-GSTexture representation

07 · Reading the pilot

What this experiment currently says

NTF transfers weakly but consistently

It improves all three still-image metrics over Vanilla 2DGS, but the magnitude is small. Cave D3 is not dominated by the same stationary synthetic sensor noise as the x64 study.

RGB-SVF is not sufficient here

Per-channel shot/read calibration does not model moving spotlight, cast shadows, clipped highlights, and residual photometric inconsistency. The Dark3R gain does not directly carry over.

Representation matters

Texture-GS strongly improves LPIPS while sacrificing pixel-aligned SSIM. The synchronized trajectory is the important test for whether that gain is stable surface detail or view-dependent artifact.