3D Vision · Gaussian Splatting · Audited Negative Results

The Rendering–Surface Gap in Gaussian Splatting

A controlled DTU Scan 24 study showing that high-fidelity novel-view rendering and accurate extracted geometry are related but distinct objectives. The artifact compares published PGSR, MILo, and TSGS baselines with project diagnostics that were accepted only when both image and surface criteria improved.

Gaussian Splatting Surface Reconstruction DTU Scan 24 Chamfer Distance Novel-View Synthesis Acceptance Gates Reproducible Audit

Ground truth, PGSR baseline, and GaugeSplat candidate renders for held-out DTU view 24

Held-out view 24: ground truth, PGSR, and the GaugeSplat diagnostic. The candidate is visually almost indistinguishable from its parent, yet the extracted-surface change is only 0.17%.

What this project established

23.480 dB

Best PSNR: local TSGS reproduction

0.0987

Best LPIPS: local TSGS reproduction

2.431 mm

Best third-party control Chamfer: MILo

3.86%

TSGS post-hoc extraction Chamfer reduction

Ownership boundary. PGSR, MILo, and TSGS are published third-party methods. The project contribution is the cross-method audit, mathematical framing, controlled probes, paired acceptance gates, and reproducible result artifact—not a new state-of-the-art renderer.

Reproduction boundary. The local TSGS run achieved the strongest image metrics, but its archived official-geometry gate failed: 2.752 mm measured Chamfer versus 0.391 mm reported by the authors. It is presented as a measured local reproduction, not a successful reproduction of the published geometry number.

Analytical framing

Let θ denote the parameters of a Gaussian scene, R(θ, c) its rendered image from camera c, and E(θ) the surface produced by a chosen extraction operator. Training primarily minimizes an image-space objective, while evaluation of geometry acts on a different object produced after extraction.

L_img(θ) = Σ_c∈Ctrain ℓ(R(θ,c), I_c)

L_geo(θ; E) = ½[d(E(θ) → S*) + d(S* → E(θ))]

Low image loss therefore does not imply low surface error: the renderer integrates color, opacity, covariance, and visibility, while extraction applies a non-equivalent depth or density rule. A candidate perturbation is retained only when it crosses a geometry threshold without violating held-out image tolerances.

accept(θ′) ⇔ ΔL_geo ≤ −τ_geo and ΔM_img within ε_img

Reporting both directional distances matters. The TSGS first-surface diagnostic lowers surface-to-data distance but raises data-to-surface distance; its average improves even though the reconstruction becomes less complete in one direction.

Third-party method control runs

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FG PSNR ↑	Chamfer ↓
PGSR	20.258	0.7210	0.2372	21.620	3.287 mm
MILo	20.565	0.7383	0.2158	20.787	2.431 mm
TSGS	23.480	0.9149	0.0987	27.574	2.752 mm

All image metrics use the same seven held-out views. Results are limited to DTU Scan 24.

Controlled project diagnostics

Diagnostic	Parent	PSNR Δ	Chamfer	Relative change	Decision
RayOT	PGSR	−0.186 dB	3.111 mm	+5.36%	Reject: image
GaugeSplat	PGSR	−0.001 dB	3.281 mm	+0.17%	Reject: geometry
TraceSplat	MILo	−0.019 dB	2.426 mm	+0.19%	Reject: geometry
TSGS first surface	TSGS	unchanged	2.646 mm	+3.86%	Post-hoc diagnostic

RayOT shows the central tradeoff most clearly: geometry improved, but PSNR, SSIM, LPIPS, and foreground PSNR all worsened. GaugeSplat and TraceSplat preserved rendering but did not move geometry enough to pass the geometry gate. VP0 also failed its scale-up criterion.

Representative held-out renders

Ground truth · held-out view 0

Local TSGS reproduction · strongest image metrics in this study; official geometry target not reproduced

MILo control · strongest measured third-party geometry score

TraceSplat diagnostic render of held-out DTU view 0

TraceSplat diagnostic · rendering preserved, geometry gate not passed

TSGS depth rendering for held-out view 24

TSGS depth at view 24 · input to the post-hoc first-surface analysis

Reproducibility and limitations

Included

Exact JSON metric snapshot used by the report
Deterministic CSV and Markdown table generator
Numerical validation and unit tests
Seven-page PDF and LaTeX source archive
Representative renders and comparisons

Not claimed

No DTU-wide or state-of-the-art superiority
No ownership of PGSR, MILo, or TSGS
No unified redistribution of third-party training code or weights
No conclusion beyond the single audited scan
No positive method claim for rejected diagnostics