Scene-Conditioned Distributional Quality Assessment for Radiance Fields

Saeed Mahmoudpoura,b, Gi-Mun Umc, Hyon-Gon Chooc, and Peter Schelkensa,b

aDepartment of Electronics and Informatics, Vrije Universiteit Brussel, Belgium
bIMEC, Kapeldreef 75, Leuven, B-3001, Belgium
cImmersive Media Research Section, Electronics and Telecommunications Research Institute (ETRI), 218 Gajeong-Ro, Yuseong-Gu, Daejeon, 34129, Republic of Korea

SCODA-QA evaluates rendered views against a scene-specific distribution of clean reference views. It is designed for radiance-field and light-field content where exact pixel-aligned references may be unavailable or unreliable.

View Code Get Started View Method

Method

Quality from a Scene Manifold

SCODA models each scene using clean reference views. A pretrained CNN extracts semantic features, and a multivariate Gaussian captures the normal distribution of high-quality observations for that scene.

A test rendering receives a fidelity score, QF, from its Mahalanobis distance to this scene model. A lightweight patch discriminator provides QR, a distortion-aware realism cue. The final score applies bounded residual fusion after quantile alignment.

The default mode treats each rendered image as one view. For light fields sampled on a regular angular grid, SCODA can split each view into an N x N patch grid, score local regions, and pool them into one view-level quality estimate.

  • Scene-conditionedUses multiple reference views rather than a single aligned image.
  • DistributionalRepresents reference content as a scene-specific feature distribution.
  • Distortion-awareCombines semantic fidelity with a trained realism signal.
Overview of the SCODA-QA method.

Results

Representative Validation

3DGS-15K QF SRCC0.8819
3DGS-15K QR SRCC0.8716
3DGS-15K Q SRCC0.8982

On the 3DGS-IEval-15K validation set, the scene-conditioned fidelity term and realism term are both strongly correlated with human opinion scores. Their fused score gives the best rank correlation while keeping the method usable without a precisely aligned full-reference image.

Score Meaning SRCC KRCC PLCC
QF Scene-manifold fidelity 0.8819 0.6938 0.8796
QR Distortion-aware realism 0.8716 0.6770 0.8703
Q Bounded fused SCODA score 0.8982 0.7154 0.8959
When to use SCODA-QA: use it when a scene has a set of clean training or reference views, but each rendered test view does not have a perfectly aligned reference image. This is common in NeRF, 3DGS, and light-field quality assessment.

Code

Run SCODA-QA

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python run_scoda.py \
  --image /path/to/test.png \
  --scene bicycle \
  --scene-ref-root /path/to/scene_train

Use --spatial-mode image for single rendered views and --spatial-mode grid --grid-size 3 for LF-style patch scoring. Cached stats are checked against the requested mode.

The public repository is available at gitlab.com/saeedmp/scoda.

Citation

Cite

@inproceedings{scodaqa2027,
  title={SCODA-QA: Scene-Conditioned Distributional Quality Assessment for Radiance Fields},
  author={Mahmoudpour, Saeed and Um, Gi-Mun and Choo, Hyon-Gon and Schelkens, Peter},
  booktitle={...},
  year={2027}
}