acceptodds
Under review as a conference paper at ICLR 2027

Beyond Coverage: Single-View Signals for Post-Hoc Uncertainty Estimation in Gaussian Splatting

Abstract

Post-hoc uncertainty estimation for Gaussian Splatting aims to predict rendering errors without modifying a trained scene representation. Existing approaches largely rely on multi-view coverage statistics, requiring training-view rendering, ground-truth comparison, and back-projection onto Gaussian primitives. We revisit where predictive uncertainty signals originate and show that much of the signal is already present in a single rendering at the novel pose. Four local appearance statistics extracted from the colour buffer are as predictive as the strongest coverage-based features, while a simple gradient-magnitude map remains competitive without any learned regressor. These appearance cues and geometry statistics are largely complementary: a representation combining both, together with scale-invariant statistics from the rendered depth buffer, improves Pearson correlation by 23%, Spearman correlation by 17%, and reduces AUSE by 14% over the strongest post-hoc baseline across 23 scenes from four datasets. This scale-invariant representation further enables a single estimator trained across scenes to transfer to unseen scenes without target-scene fitting. We also identify monotone view-dependent feature drift as a source of evaluation bias, and show that an evidential head on within-view rank representations, which removes this drift by construction, adds calibrated predictive intervals at no cost in ranking accuracy. These findings suggest that reliable post-hoc uncertainty in Gaussian Splatting depends not only on scene observation history but also on signals intrinsic to the rendered view itself.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty Quantification, Gaussian Splatting, Post-hoc Uncertainty Estimation, Evidential Regression, Distribution Shift, Invariant Representation, Cross-Scene Generalization, Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6vWxOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies post-hoc uncertainty estimation for Gaussian Splatting: predicting rendering error for a novel view without altering the trained scene representation. Prior work relies on multi-view coverage statistics obtained by rendering and back-projecting training views. This paper shows that simple appearance and depth statistics computed from a single novel-view rendering already carry much of that predictive signal, and that combining the two sources, together with a rank-based representation robust to view-dependent feature drift, improves on the strongest published post-hoc baseline and enables a single estimator that transfers across scenes without per-scene fitting.

Critical strengths

S1. The diagnosis that monotone, view-dependent feature drift can bias the evaluation of post-hoc uncertainty estimators, and the resulting design choice to represent features by their within-view rank rather than their raw value, is a small but genuine and non-obvious contribution, even though the underlying invariance property itself is a classical result.

S2. The experimental methodology is generally careful, in particular the controlled ablation (Table 6) that isolates the contribution of the regressor architecture from the contribution of the input features under identical inputs — exactly the kind of disentangling ablation that would strengthen comparisons of this sort.

S3. The paper is easy to follow. A strength, which carries a new meaning and becomes noteworthy in these days of papers densified by AI until the information content per sentence grows beyond readability.

Critical weaknesses

W1. The paper's contribution is too incremental for a venue like ICLR. Most of the paper's experimental structure already mirrors that of the embedded baseline (PRIMU) itself: the main uncertainty-estimation comparison (same three datasets, same AUSE/Pearson metrics), the feature-combination ablation, and the cross-scene generalization experiment all closely follow PRIMU's own published evaluation design, with PRIMU's released feature family simply extended by new inputs. The genuinely new empirical content is narrower than the full experimental section suggests: it rests on the new feature families themselves and on the rank-invariant evidential extension, while the surrounding experimental scaffolding is largely reused. Conversely, two of PRIMU's own experiments — a foreground/background breakdown and a downstream active-view- selection evaluation — are not carried over, which relates to the scope concern below.

W2. The scope is narrow for a venue like ICLR: the paper evaluates purely at the level of uncertainty-estimation metrics (Pearson, Spearman, AUSE) and never demonstrates that the improved numbers translate into a benefit for a downstream task (e.g., active view selection, anomaly detection, or scene change detection), even though such downstream validation is standard in closely related work on the same problem, see FisherRF, PRIMU and also Galappaththige et al. (ECCV 2026).

W3. Two aspects of the paper's framing overstate what is actually shown. First, the central contrast ("existing methods need multi-view coverage, we show one rendering suffices") conflates a one-time, inexpensive per-scene setup cost with an untested assumption about the baseline's requirements, since both approaches (the proposed one and PRIMU) need only a single rendering to evaluate a new pose once that one-time setup is done. The core reason is that both works build on a pixel-wise regressor. One image contains enough pixels to learn the task. Second, labeling the two components of the evidential decomposition as "aleatoric" and "epistemic" uncertainty is adopted to 3DGS without discussing whether this terminology is meaningful for a static, deterministic scene representation, where the usual justifications/assumptions for this distinction do not obviously apply. The use of "aleatoric" and "epistemic" should be reconsidered.

W4. The paper omits any related-work discussion or empirical comparison with Galappaththige et al. (ECCV 2026), which addresses the same problem (primitive-/ rendering-based uncertainty estimation for Gaussian Splatting) on the same benchmark (Mip-NeRF 360). On the one metric that is directly comparable across the two papers' publicly reported numbers (AUSE on the L1 target, Mip-NeRF 360), the two methods appear to perform essentially on par. A preprint of this work was available many months before the ICLR submission deadline. However, since the ECCV release is fairly new, this concern is mostly to be taken into account for future submissions.

Questions for authors

Q1. How does this work relate to Galappaththige et al. (ECCV 2026), which addresses the same problem on the same benchmark and was public in terms of a preprint well before this submission's deadline? A direct empirical comparison, in particular on AUSE under the L1 target where the two methods' publicly reported numbers appear comparable, is currently missing.

3 comments

Dear Reviewer 6vWx,

Thank you for a careful review. Your strengths S1–S3 describe what we hoped the paper would offer, and your weaknesses are specific enough that we could act on each.

Revision status. We have now uploaded a revised PDF. We have clarified the eight-map configuration and transfer scope, corrected Table 1 LLFF PSNR to 23.70, updated the FisherRF reference to ECCV 2024, and reconciled Figure 3 and its caption with the 18-scene means reported in the text. The additional experimental results are reported below.

Our response in three points.

  • No target-scene data needed. A global estimator on eight single-rendering maps, with no coverage maps and no training views, ground truth or fitting on the target scene, reaches Pearson 0.309 in leave-one-scene-out over 23 scenes (coverage-only global: 0.292; coverage-only per-scene fitting: 0.311).
  • Sparse views. With 12 or 24 training views, the full representation beats the coverage-only baseline in all 6 configurations, by 0.082 Pearson on average.
  • Better pixel-error ranking. Discarding the most uncertain pixels gives +0.48 to +1.12 dB higher PSNR on the retained pixels than coverage-only. This is retained-pixel PSNR, a measure of ranking quality; the rendering itself is unchanged.

W1 (the contribution is too incremental). You are right that the datasets, metrics, ablation structure and regressor follow PRIMU. As you note in S2, this is what makes the attribution clean: everything else is held fixed. We do not claim a new regressor. The contribution is a set of findings that, to our knowledge, PRIMU does not report: a Sobel gradient map used directly, with no training, reaches Pearson 0.292 against 0.311 for the full 13-map coverage pipeline (23 scenes, Table 2); appearance and coverage contribute about equally (Shapley split 49.7% / 50.3%) and are weakly related (prediction-level r× = 0.29); a single global regressor trained on the other 22 scenes matches per-scene fitting on average (Δ = +0.002 Pearson, p = 0.93; this is not an equivalence test), whereas PRIMU trains per scene; and same-view pixel splits raise gradient-boosting Pearson from 0.378 to 0.637, so the evaluation split alone changes conclusions about feature utility. This reply adds the coverage-free estimator (W3), sparse views, and zero-shot transfer to 2DGS renders (Pearson 0.310 on four scenes against 0.333 on 3DGS). We agree with S1 that rank invariance is classical; what we contribute is the diagnosis that view-dependent drift can bias this evaluation, and the data on how large the split effect is.

W2 (no downstream task). We added a deployment-style evaluation on all 23 scenes. For each view we discard the top-f fraction of pixels ranked by each estimator's predicted uncertainty and compute PSNR on the retained pixels only (300k sampled pixels per view). The splatting is unchanged by every estimator (renders are bit-identical, Appendix B), so this measures ranking quality, not rendering quality. Mean retained-pixel PSNR (dB):

f removedcoverage-onlyfull ΦΦapp∪Φgeo
0.128.0128.4927.89
0.329.9030.7529.72
0.531.4632.5831.25

The full representation beats coverage-only at every f (+0.48 dB at 10%, +1.12 dB at 50%; paired p < 0.001) and on every dataset; Φapp∪Φgeo alone is on par with coverage-only. Ranking whole views by mean predicted uncertainty does not differ significantly between estimators. In a preliminary active-view-selection run (4 scenes, 3 seeds) our selectors did not improve final PSNR over random selection (19.4 and 19.6 dB against 20.85 dB), so we make no claim for view selection; AVS, anomaly detection and change detection will be listed as open.

W3 (framing). Both points are fair; we have revised the corresponding wording.

"One rendering suffices." You are right that per-pose inference needs one rendering for both methods, and that both use a pixel-wise regressor. That sharpens the real question: what must the target scene supply? Coverage maps need every training view rendered, compared with ground truth and back-projected (60–120 s per scene, Table 4); the eight single-rendering maps need only the rendering being evaluated (about 4 s). The revised Introduction now explicitly states the eight-map configuration and its target-scene input requirements. We also ran the direct test, a global estimator on the eight single-rendering maps only (leave-one-scene-out, 23 folds):

InputsGlobal Pearson / AUSERequires target-scene training views + GT
Coverage only (13 maps)0.292 / 0.388Yes
Single-rendering maps (8)0.309 / 0.365No
All (21 maps)0.383 / 0.316Yes (coverage part)

The coverage-free estimator reaches Pearson 0.309 compared with 0.292 for the coverage-only global estimator, and is within 0.002 of coverage-only per-scene fitting (0.311). Given a frozen target-scene GS model, it requires no target-scene training images, ground-truth images or target-scene regressor fitting. We do not claim these differences are significant, and the full combination performs best among these configurations. Global and per-scene fits use different pixel budgets, as stated in the paper.

Aleatoric/epistemic. We agree that the classical justification does not obviously apply to a deterministic static scene. We have renamed the terms "noise term E[σ²]" and "evidence term Var[μ]" and explicitly stated this naming convention, without claiming that the two components identify distinct physical sources of uncertainty.

W4 and Q1 (Galappaththige et al.). Thank you; we have added this work to Related Work and clarified its relationship to our approach. On Mip-NeRF 360 they report AUSE(L1) 0.328 and Pearson(L1) 0.369; our full representation gets 0.324 and 0.299. AUSE is on par and their Pearson is higher. As we understand it, the protocols differ (they fit on training-view residuals; we train on one hold-out view and evaluate on disjoint views), so we report the published numbers side by side without claiming a win. A like-for-like comparison would require running their code under our protocol, which we have not done.

Overall assessment. We have corrected the framing and terminology and added the missing related-work discussion. We have also conducted the additional evaluations reported above. Our claims are scoped to per-pixel RGB error prediction on static scenes. If a specific soundness concern remains, we would welcome the opportunity to address it directly. We would be grateful if you could reconsider the soundness and overall scores in light of these clarifications, additional results and the revised manuscript.

Dear Reviewer 6vWx,

We would like to briefly follow up on our updated responses. The revised PDF clarifies the target-scene input requirements, revises the evidential terminology, and adds the missing related-work discussion. Our responses also report the additional sparse-view, 2DGS, and pixel-error-ranking evaluations.

If these revisions and results alleviate your concerns, we would be grateful if you could consider raising your soundness and overall ratings. We would be happy to clarify any remaining issues.

Thank you for your time and consideration.

Reviewer_bz3yOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes eight maps to include appearance and geometry feature families as well as the coverage feature family for post hoc uncertainty estimation in vanilla 3D GS. To figure out where the signal for post-hoc uncertainty estimation come from and what each source contributes, they design a unified view-disjoint protocol. With the unified representation including coverage, appearance and geometry 21 maps, the proposed model improves Pearson and Spearman correlation and AUSE compared with the baselines.

Critical strengths

The main results are informative and clear. The pipeline protocol [P] and offline protocol [O] carries the per-scene and cross-scene experiment individually and validate the model best baselines on both cases. The ablation studies are also clear which demonstrate the effectiveness of different combination of coverage, appearance, and geometry maps.

Critical weaknesses
  1. The model seems to only support the 3D GS under the static scene and be hard to extend to 4D GS directly without fundamental changes. Is this the limitation to the model?
  2. Compared with the baseline PRIMU, the method includes appearance and geometry feature families, but the 8 maps looks like additional addons of the PRIMU regressor, which may be considered as an extension work of PRIMU.
Questions for authors

Besides of the Weakness section, I have questions as below:

  1. In the experiment section, 23 scenes are selected from 4 different datasets. Could authors explain the scene selection criterion and if the model works on other scenes?
  2. In the experiments, the paper discusses per-scene fitting and global leave-one-scene-out fitting, but what happen to leave-one-dataset-out fitting? Is its result close to the global fitting?
  3. Does the full representation help the downstream tasks such as post optimization, active perception, and uncertainty-guided inpainting? Could the author give some downstream demos?
  4. Besides, I have two extended questions: What happens to the sparse-view setting? And does the model work on different versions of 3DGS including per-scene optimization based methods and feedforward methods?
1 comment

Dear Reviewer bz3y,

Thank you for the review and the concrete questions.

Revision status. We cannot upload a revised PDF at this time. Changes marked "will" will be made in the revised version, and all new results are given here. We will also correct three errors: Table 1 LLFF PSNR should read 23.70; the FisherRF reference should cite ECCV 2024; Figure 3's bars do not match the numbers in the text.

Our response in three points.

  • No target-scene data needed. A global estimator on eight single-rendering maps, with no coverage maps and no target-scene views, ground truth or fitting, reaches Pearson 0.309 in leave-one-scene-out over 23 scenes (coverage-only global: 0.292; coverage-only per-scene: 0.311).
  • Sparse views. The full representation beats coverage-only in all 6 sparse-view configurations, by 0.082 Pearson on average.
  • Better pixel-error ranking. Discarding the most uncertain pixels gives +0.48 to +1.12 dB higher PSNR on the retained pixels than coverage-only; this is retained-pixel PSNR, not an improvement of the rendering.

W1 (static 3DGS only; 4D). Correct: the method is evaluated on static scenes and we have not tested dynamic or 4D representations. We will state this as a limitation.

W2 (an extension of PRIMU). The regressor is PRIMU's by design, to keep feature-family comparisons interpretable. The contribution is the accounting of where the signal comes from and the evaluation protocol: a Sobel map used directly, with no training, reaches Pearson 0.292 against 0.311 for the full coverage pipeline (23 scenes); appearance and coverage contribute about equally (Shapley split 49.7% / 50.3%); and same-view pixel splits inflate gradient-boosting Pearson from 0.378 to 0.637. The transfer results below show the maps are not a trivial add-on.

Q1 (scene selection). The 23 scenes are all nine Mip-NeRF 360 scenes, all eight LLFF scenes, all four Light Field scenes, and two Tanks and Temples scenes (truck, train). We did not tune the set. Beyond these scenes, we also trained 2DGS on four of them (see Q4).

Q2 (leave-one-dataset-out). Pearson on each held-out dataset, global estimator trained on the other datasets:

held outcoverage (13)single-rendering maps (8)full Φ (21)per-scene fit, full Φ
Mip-NeRF 360 (9)0.2150.2600.3060.310
LLFF (8)0.1330.3280.2600.444
Light Field (4)0.2740.3070.3360.497

For Mip-NeRF 360 the full global estimator matches the per-scene fit. For LLFF and Light Field dataset shift costs accuracy relative to per-scene fitting; coverage-only suffers most, while the eight single-rendering maps are the most robust single arm on LLFF. The leave-one-scene-out result in the paper (global matches per-scene, Δ = +0.002, p = 0.93, not an equivalence test) therefore holds when scenes of the same dataset are in training, and we will add this caveat. (Tanks and Temples has two scenes and is reported separately.)

Q3 (downstream tasks). We added a deployment-style evaluation on all 23 scenes: for each view we discard the top-f fraction of pixels ranked by predicted uncertainty and compute PSNR on the retained pixels (300k sampled pixels per view). The splatting is unchanged by every estimator, so this measures ranking quality. The full representation beats coverage-only at every f (+0.48 dB at 10%, +1.12 dB at 50%; paired p < 0.001). In a preliminary active-view-selection run (4 scenes, 3 seeds) our selectors did not improve final PSNR over random selection (19.4 and 19.6 dB against 20.85 dB), so we make no claim for view selection. Uncertainty-guided inpainting and post-optimization were not done.

Q4 (sparse views, other versions, feed-forward). Sparse views: with 12 or 24 training views (bicycle, garden, bonsai, room), the full representation beats coverage-only in all six configurations (Pearson 0.319 vs 0.237; AUSE 0.354 vs 0.442; p < 0.001 over scenes). Another representation: the eight-map estimator trained on 3DGS only gives Pearson 0.310 on 2DGS renders of four scenes against 0.333 on 3DGS, zero-shot (the coverage pipeline does not support 2DGS, so there is no coverage comparison). Mip-Splatting and feed-forward models are untested and will be listed as limitations.

Presentation. In the revision the introduction will define the three feature families and the protocols [P] and [O] before use, Figure 1 will be larger, and §3.2 will move to the appendix. If other parts were hard to follow, please tell us.

We hope these additions address your questions, and we would be grateful if you would take them into account in your final assessment.

Reviewer_NQyyOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies post-hoc uncertainty estimation for 3D Gaussian Splatting: predicting the per-pixel rendering error of a frozen, already-trained model. The authors focus on how much of that predictive signal is already in the single rendered view at the novel pose, and develop a single regressor that can generalize to novel scenes. Experimental results are validated on standard benchmarks in this task.

Critical strengths
  1. The idea is new for uncertainty estimation in Gaussian Splatting. This paper shows that a Sobel filter on the rendered image, with no training, achieves a Pearson correlation of within 0.02 of a pipeline that re-renders and back-projects every training view, which is an important result. Future post-hoc uncertainty papers should report this baseline, and I think that alone has value.

  2. The experimental design is comprehensive. The authors verify their method by improvements on 23/23 scenes for both Pearson and AUSE, and the metrics reported and scene tested are complete to present the ablation.

  3. The finding on evaluation splits is practically important. The reported Pearson coefficient drops from 0.637 to 0.378 for gradient boosting when evaluation moves from same-view pixels to disjoint views. This is a good warning for anyone benchmarking in this area.

Critical weaknesses
  1. A limitation is about scope. Only vanilla 3DGS is tested, and it would help to show the findings hold for other variants such as Mip-Splatting or 2DGS, or for sparse-view models. On Light Field, the reproduced splatting is about 5 dB below the published one (28.86 vs 34.18 PSNR), so baseline numbers there aren’t directly comparable with the literature. FisherRF also shows negative Pearson on all three datasets in Table 1, which seems surprising enough to warrant a short explanation or a check of the implementation.

  2. Some inconsistency exist in the paper: a. The LLFF PSNR doesn’t match. Table 1 lists LLFF PSNR as 24.70, but Appendix C says “Our LLFF means are 23.70 / 0.792 / 0.137”. The per-scene values in Table 5 average to 23.70, so Table 1 has a typo. b. The FisherRF venue doesn’t match. Table 1 says “ECCV’24”, but the reference list cites it as an arXiv 2023 preprint.

  3. The performance of this uncertainty estimation is only provided for per-pixel L1 RGB error. The author can provide results such as using it in Active Reconstruction, such as done in FisherRF.

Questions for authors
  1. Why does FisherRF give negative correlation with rendering error in your runs? Did PRIMU itself use same-view evaluation? In other words, does the split finding change any previously published conclusion, or is it a general warning?

  2. How does the comparison between appearance and coverage change in sparse-view or extrapolated-view settings, where coverage should vary much more across the image?

1 comment

Dear Reviewer NQyy,

Thank you for the careful review, and for singling out the Sobel baseline and the evaluation-split finding as strengths.

Revision status. We have now uploaded a revised PDF. The manuscript corrections are summarized under W2, and the additional experimental results are reported below.

Our response in three points.

  • No target-scene data needed. A global estimator on eight single-rendering maps, with no coverage maps and no target-scene views, ground truth or fitting, reaches Pearson 0.309 in leave-one-scene-out over 23 scenes (coverage-only global: 0.292; coverage-only per-scene: 0.311).
  • Sparse views. The full representation beats coverage-only in all 6 sparse-view configurations, by 0.082 Pearson on average.
  • Better pixel-error ranking. Discarding the most uncertain pixels gives +0.48 to +1.12 dB higher PSNR on the retained pixels than coverage-only; this is retained-pixel PSNR, not an improvement of the rendering.

W1 (scope; Light Field PSNR; FisherRF).

Other variants and sparse views (see also Q2). As a first step beyond vanilla 3DGS we trained 2DGS on garden, bicycle, bonsai and room and applied the global eight-map estimator, trained only on 3DGS features of other scenes, zero-shot: Pearson 0.310 on 2DGS against 0.333 on 3DGS for the same scenes (AUSE 0.326 vs 0.308). The coverage pipeline does not support 2DGS, so there is no coverage comparison; four scenes only, and we make no claim for other variants. Mip-Splatting and feed-forward models are untested, and we will say so in the limitations.

Light Field PSNR. The 28.86 vs 34.18 dB gap is disclosed in Appendix C and remains unexplained after matching resolution, extending training and masking borders. Light Field comparisons use our own PRIMU reproduction on the same splatting, so baseline and method are compared fairly, and the gain is stable across the settings we ran (+0.085, +0.074, +0.077 Pearson).

FisherRF negative Pearson. We re-checked sign convention and normalization. The map is built from the diagonal of the Fisher information accumulated per primitive, so a high value means a strongly constrained primitive; it measures constraint strength, not error size. A negative correlation with error is the expected reading, and the published one: PRIMU's Table 1 reports FisherRF Pearson −0.13 (Light Field), −0.054 (Mip-NeRF 360) and −0.088 (LLFF), and our Light Field run gives −0.144. We found no implementation error and will add this explanation to Appendix C.

W2 (inconsistencies). Both are our errors, and we will correct them: Table 1 LLFF PSNR should read 23.70 (Appendix C and Table 5 are correct), and the FisherRF reference should cite ECCV 2024. We also found that Figure 3's bars (0.464→0.161 and 0.664→0.386) do not match the text (0.508→0.138 and 0.637→0.378); we will make the figure match and state which scenes each number covers.

W3 (only per-pixel L1 error; active reconstruction). We added a deployment-style evaluation on all 23 scenes: for each view we discard the top-f fraction of pixels ranked by predicted uncertainty and compute PSNR on the retained pixels (300k sampled pixels per view). The splatting is unchanged by every estimator, so this measures ranking quality. The full representation beats coverage-only at every f (+0.48 dB at 10%, +1.12 dB at 50%; paired p < 0.001). In a preliminary active-view-selection run (4 scenes, 3 seeds) our selectors did not improve final PSNR over random selection (19.4 and 19.6 dB against 20.85 dB), so we make no claim for active reconstruction.

Q1 (FisherRF; did PRIMU use same-view evaluation?). FisherRF: see W1. PRIMU trains on one hold-out view and evaluates on the remaining views, which is view-disjoint, and our reproduction matches its numbers. Our same-view finding is a general warning about comparing estimators under different splits; it does not overturn PRIMU's conclusions.

Q2 (sparse or extrapolated views: appearance vs coverage). We retrained 3DGS with 12 or 24 training views (test split unchanged) on bicycle (12 and 24 views), garden (12 and 24), bonsai (24) and room (24), and re-fitted per-scene regressors on the same frozen splattings. Pearson / AUSE, mean over the six configurations:

coverage-only (13)appearance (4)single-rendering maps (8)full Φ (21)
0.237 / 0.4420.264 / 0.4140.280 / 0.3780.319 / 0.354

The full representation beats coverage-only in all six configurations and on all four scenes (+0.082 Pearson on average; p < 0.001, paired over scenes with configurations averaged within scene), and the eight single-rendering maps alone beat coverage-only in 5 of 6. So coverage does not become more important with sparse views: the single-rendering maps stay competitive. Extrapolated views were not tested.

We hope these results and corrections address your concerns, and we would be grateful if you would consider them in your final assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.