acceptodds
Under review as a conference paper at ICLR 2027

PCAsplat: Gaussian Splatting with Local PCA Regularization

Abstract

Gaussian splatting has emerged as a flexible representation for 3D reconstruction from posed images. However, existing methods are optimized primarily using rasterization-based losses, which supervise a splat only when it contributes to sampled camera rays. Gaussians that are occluded or contribute little to the sampled view therefore receive weak or no geometric gradients and may drift away from the underlying surface, producing undesired floaters. We introduce PCAsplat, a geometry-aware regularization framework for Gaussian splatting based on differentiable local principal component analysis (PCA). Our PCA regularizer acts directly on neighborhoods of Gaussian centers and can therefore update Gaussians that do not contribute to the current training view. We regularize the PCA eigenvalues to encourage Gaussians to move to the underlying surface with isotropic tangent-plane coverage. We also align each Gaussian normal with the PCA-estimated neighborhood normal to enforce consistent orientation. Experiments on DTU, Tanks and Temples, and NeRF Synthetic show that the splats produced by PCAsplat better approximate samples of the reference surface while substantially reducing undesired floaters. These surface-aligned splats enable downstream geometry-processing tasks, including point cloud segmentation, and direct Poisson reconstruction. Additionally, PCAsplat remains competitive under conventional novel view synthesis and mesh extraction tasks. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Surface Reconstruction, PCA, Point Cloud, inverse rendering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_A3ZJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PCAsplat, a geometry-aware regularization framework for Gaussian splatting based on differentiable local principal component analysis (PCA). The main motivation is to address poorly localized Gaussian primitives and floaters that may receive insufficient geometric supervision from rendering-based losses.

The proposed method constructs local neighborhoods of Gaussian centers and introduces an eigenspectrum regularization loss to encourage local planarity and isotropic tangent-plane coverage. An additional normal alignment loss encourages consistency between Gaussian orientations and PCA-estimated surface normals. The framework is implemented on top of an existing Gaussian splatting pipeline.

The authors evaluate their method on DTU, Tanks and Temples, NeRF Synthetic, and ScanNet++, reporting improvements in point-cloud surface accuracy, instance segmentation, and Poisson surface reconstruction, while maintaining competitive novel-view synthesis and mesh reconstruction performance.

Critical strengths
  1. Effective geometric regularization with convincing improvements in point-cloud quality. The paper addresses a meaningful limitation of rendering-based Gaussian Splatting methods: insufficient geometric supervision for Gaussians with weak rendering contributions. Directly regularizing local Gaussian neighborhoods is a reasonable approach that can complement existing rendering objectives. The quantitative results are encouraging, particularly the improvement in mean Chamfer Distance on DTU from 0.662 mm for RaDe-GS to 0.519 mm. The qualitative comparisons also suggest better surface localization and fewer floating artifacts.

  2. Useful downstream evaluations beyond conventional rendering metrics. I appreciate that the paper evaluates Gaussian centers as geometric surface samples rather than focusing exclusively on novel view synthesis. The segmentation and Poisson reconstruction experiments help demonstrate the potential practical value of improved Gaussian localization. In particular, the reported improvements in ScanNet++ segmentation are substantial. The additional experiments integrating PCA regularization into different Gaussian Splatting methods also provide useful evidence of its potential generality.

Critical weaknesses
  1. The technical novelty and the individual contributions of the proposed regularization terms are not sufficiently established. While the overall formulation is reasonable, local PCA and covariance eigenspectrum analysis are well-established techniques in point-cloud processing. The proposed method appears to be a relatively straightforward application of these classical geometric priors to Gaussian centers. I would like to see stronger evidence explaining what makes the specific formulation particularly effective compared with simpler local planarity or neighborhood regularization alternatives.

More importantly, the current ablation study does not fully isolate the contributions of the different terms. In Table 4, the PCA loss accounts for almost all the improvement, whereas adding the normal alignment loss only changes the Chamfer Distance from 0.520 to 0.519. The planarity and tangent-plane isotropy terms are also not evaluated separately. Consequently, it remains unclear which components are essential and whether the full proposed formulation is justified by the experimental evidence.

  1. The practical limitations and trade-offs of the local PCA regularization require more thorough analysis. The method encourages locally planar Gaussian neighborhoods, but local planarity alone does not necessarily imply alignment with the correct scene surface. For example, isolated clusters of floating Gaussians could potentially form locally coherent structures while remaining geometrically incorrect. The paper acknowledges that an additional depth-based step is required for distant floaters, but its precise role and impact are not sufficiently explained. This makes it difficult to determine how much of the improvement can be attributed to the proposed PCA regularization itself.

Furthermore, the computational overhead is non-negligible. On Tanks and Temples, the reported training time increases from 18.3 minutes for RaDe-GS to 53.0 minutes for PCAsplat. The novel view synthesis results in Table 8 also indicate some degradation compared with RaDe-GS, particularly on outdoor scenes. A more explicit analysis of these trade-offs would help establish the practical benefits and applicability of the method.

Questions for authors
  1. Contribution of individual loss terms. Could the authors provide separate ablations for the planarity term, tangent-plane isotropy term, and normal alignment term? Given the small additional improvement from normal alignment in Table 4, it would be useful to clarify its importance, particularly for downstream tasks relying on surface normals. A comparison with a simpler neighborhood planarity regularizer would also help establish the necessity of the proposed formulation.

  2. Handling distant floaters and neighborhood sensitivity. The conclusion mentions a depth-based step for distant floaters. Could the authors clarify whether this step is part of training or post-processing, and quantify its contribution? Additionally, how sensitive is the method to the number of nearest neighbors (K), especially near sharp boundaries, thin structures, or regions with nonuniform Gaussian density?

  3. Computational cost and rendering trade-offs. Could the authors provide more detailed analysis of the runtime overhead introduced by neighborhood construction and differentiable PCA? It would also be helpful to examine the trade-off between surface reconstruction accuracy and novel view synthesis quality as the PCA regularization strength varies.

  4. Downstream segmentation evaluation. For the ScanNet++ experiments, how are differences in point-cloud size and sampling density handled before feeding the reconstructed points into Mask3D? Since these factors can affect segmentation performance, additional clarification would help establish how much of the reported improvement comes specifically from better surface alignment.

1 comment
AuthorsOpenReview ↗

Official Comment

Thank you for the feedback. We added simpler baselines, loss ablations, neighborhood-size experiments, same-GPU timings, and segmentation comparisons with equal point counts.

(W1) “PCA and covariance eigenspectrum analysis are well-established techniques” and “what makes the specific formulation particularly effective compared with simpler local planarity or neighborhood regularization alternatives.”

Classical PCA-based simplification operates on an existing point cloud (Pauly et al., 2002). Our contribution is using differentiable PCA to encourage local planarity while learning a cleaner Gaussian cloud from images. Optimized jointly with rendering losses, it directly constrains Gaussian neighborhoods, including Gaussians with weak rendering contributions.

We added centroid smoothing as a simpler neighborhood baseline, with scheduled weights and detached neighborhood means (Appendix E.2):

$\mathcal{L}_{c} = \frac{1}{N}\sum_i \lVert x_i - \bar{\mu}_i \rVert^2$.

On 15 DTU scenes, PCAsplat achieves 0.519 mm CD versus 0.523 mm for centroid smoothing (Tab. 12). Their training times on RTX PRO 4500 are 4.98 and approximately 10 min, respectively. PCA provides tangent-plane isotropy and local normals, with differentiable statistics supporting future work on geometric regularizers.

(W2 / Q1) “Could the authors provide separate ablations for the planarity term, tangent-plane isotropy term, and normal alignment term?”

Tab. 11 presents the ablation:

ConfigurationCD (mm)
Baseline0.655
Normal only0.654
Planarity only0.533
Planarity + isotropy0.520
Planarity + isotropy + normal0.519

Regarding the isotropy-only experiment, Gaussians spread without sufficient geometric constraints, and optimization did not converge. Planarity loss accounts for most of the CD improvement; isotropy promotes splat coverage and isotropic neighborhoods (Fig. 3). The normal term aligns PCA normals with the Gaussian normals; however, Gaussian-center CD does not directly measure normal orientation. Fig. 19 shows more coherent normal directions in the DTU apple example. We use the centers and normals for Poisson reconstruction (Fig. 9).

(Q2) “how sensitive is the method to the number of nearest neighbors (K)”

Tab. 13 reports DTU results:

KGaussian CD (mm)Mean GaussiansMesh CD (mm)
100.593398,9380.511
200.519349,2280.471
400.494365,5760.531

Increasing K from 20 to 40 improves the Gaussian CD but increases the TSDF mesh CD. K=20 remains our default; the results suggest choosing K according to the target application. These results illustrate the trade-off across the tested settings but do not isolate behavior near sharp boundaries, thin structures, or regions with nonuniform density.

(Q2) “detailed analysis of the runtime overhead introduced by neighborhood construction and differentiable PCA?”

New same-GPU DTU measurements, with GPU memory capped at 24 GB, quantify the training overhead (Appendix E.4, Tab. 14):

GPURaDe-GSPCAsplat
RTX PRO 4500 Blackwell3.38 min4.98 min
RTX A550010.75 min22.26 min

The RTX 4500 figures measure the training loop (without logging). The overhead varies across the hardware. And we leave for future to optimize our KNN based batched PCA regularization.

(Q2) “examine the trade-off between surface reconstruction accuracy and novel view synthesis quality as the PCA regularization strength varies.”

We could not reproduce RaDe-GS’s published scores. See Tab. 9 and Reviewer RxEG
answer W5 which reports the published results and our reproduction showing a gap of only 0.14 PSNR in outdoor scenes and 0.31 in indoor scenes and similar LPIPS and SSIM. Fig. 16 show rendering comparison over two indoor and two outdoor scenes, visually showin almost no difference.

(Q2) “The conclusion mentions a depth-based step for distant floaters”
This was left from an earlier draft and was removed.

(Q3) “For the ScanNet++ experiments, how are differences in point-cloud size and sampling density handled before feeding the reconstructed points into Mask3D?”

All methods use the same ScanNet++ scenes, frozen Mask3D, and evaluation protocol. Tab. 5 reports results using the original point counts. To control point count, we added comparisons of PCAsplat and RaDe-GS centers, depth-derived points and TSDF mesh samples, using 732k points per cloud, the same eight ScanNet++ scenes, and the same frozen Mask3D model. The version show additional results in Tab. 10 which are discussed both in the paper and in Reviewer RxEG (W1). Mesh and point samples are distributed uniformly over the extracted surfaces which are later passed trough Mask3D original inference chunking for equal density. The advantage of PCAsplat's centers persists after matching point count or changing media.

Reviewer_jy1bOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper adds a local PCA regularizer to Gaussian splatting. It constrains local thickness, plane distribution, and normal. The results show that Gaussian centers are better aligned with scene surfaces, with fewer floaters. This also helps downstream tasks such as 3D instance segmentation and Poisson surface reconstruction. The final rendering results in supplementary video are competitive, but the gains on complex scenes are limited.

Critical strengths
  1. The paper is easy to follow. The problem and the motivation for using local PCA are clear. The constraints are also intuitive, and it is easy to understand what each one is trying to do.
  2. I like this idea. Making Gaussians closer to the surface is useful for reducing floaters and for downstream tasks. Local PCA is a reasonable way to do this. The method does have some effect: in the video, the rendering looks smoother after removing floaters.
Critical weaknesses
  1. The rendering looks smoother in several examples in the supplementary video after removing floaters. However, the overall metrics in Table 8 do not show the same improvement. I would like the authors to explain why the video looks better but the metrics do not improve.
  2. If the focus is geometry, I would like to see more mesh results. Figures 4, 5, and 6 mostly show Gaussian centers and cross-sections. The mesh comparisons are mainly on DTU. More examples on Mip-NeRF360 and Tanks and Temples would be helpful, especially direct comparisons with PGSR.
  3. The results in Tables 7 and 8 show limited gains on complex datasets. The mesh results on Tanks and Temples are similar to existing methods, and there is no clear advantage in rendering on Mip-NeRF360. The Gaussians look cleaner, but the final meshes and rendering do not improve as much. I would like to understand why.
Questions for authors
  1. I like the idea, and the video results look good. But in Tables 7 and 8, the advantages in mesh quality and rendering are not very clear. What do the authors consider the main advantage of this work?
  2. Could the authors add more results to show this more clearly? I would like to see rendering comparisons and metrics with and without the regularizer, as well as more mesh results on complex scenes. A direct comparison with PGSR on the same scenes and from the same viewpoints would also be helpful.
2 comments
AuthorsOpenReview ↗

Official Response by Authors

We thank the reviewer for the positive feedback. Below, we clarify the rendering results and provide additional mesh comparisons.

(W1) “I would like the authors to explain why the video looks better but the metrics do not improve.”

Thank you for noting the smoother rendering in the supp. videos. PSNR, SSIM, and LPIPS evaluate fixed test views within the training cameras domain, where occluded or weakly contributing floaters can have little effect on the scores. These floaters can become visible during camera motion, so reducing them can improve video quality without substantially changing average image metrics. This motivates evaluating Gaussian centers directly (Tabs. 1-3, Fig. 11). Additionally, we tried reproducing the metrics with RaDe-GS, and the refactored Tab. 9 shows very similar photometric scores; see also (W5) of RxEG.

(W2) “More examples on Mip-NeRF360 and T&T would be helpful, especially direct comparisons with PGSR.”

Fig. 15 adds direct comparisons with PGSR and RaDe-GS on Ignatius and Truck from identical viewpoints. PCAsplat preserves the statue and vehicle structures and avoids the large foreground artifact in PGSR's Ignatius mesh; its quality is broadly comparable to RaDe-GS. PCAsplat and RaDe-GS use identical TSDF settings, while PGSR uses its official extraction pipeline.

Also, Figs. 17-18 compare raw NeRF Synthetic meshes. Without connected-component filtering, mean F1@5mm improves from 0.6199 to 0.6273, while the median number of floating components outside the 5mm reference-surface margin falls from 693.5 to 44.5 (93.6% fewer). Filtering can remove these artifacts but also discards valid instruments in Drums and foliage in Ficus. Evaluating raw meshes exposes differences that post-processing can hide.

(W3) “The Gaussians look cleaner, but the final meshes and rendering do not improve as much. I would like to understand why.”

TSDF meshes are fused from rendered depth, and image metrics evaluate rendered views. Occluded or weakly contributing floaters can have little effect on either, so reducing them can improve Gaussian geometry without substantially changing these scores. Connected-component filtering (standard post-processing for evaluating TSDF meshes) drastically removes common mesh artifacts generated by floaters (Figs. 17-18).

(Q1) “What do the authors consider the main advantage of this work?”

PCAsplat produces a cleaner Gaussian cloud by reducing floaters that degrade surface reconstruction and rendering. Our regularization acts during training, including on Gaussians weakly constrained by rendering losses. The centers support segmentation and, together with their normals, provide oriented point clouds for Poisson reconstruction without depth rendering or mesh extraction. The Gaussians can also be cropped and reused with less manual cleanup.

(Q2) “Could the authors add more results to show this more clearly?”

We have added rendering comparisons on four Mip-NeRF 360 scenes (Kitchen, Room, Flowers, and Treehill) using identical viewpoints (Figure 16), together with reproduced RaDe-GS metrics (Table 9), qualitatively confirming similar photometric rendering quality. Also, we provide direct comparisons with PGSR and RaDe-GS on the Ignatius and Truck meshes (Figure 15), as well as raw, unprocessed mesh comparisons on NeRF Synthetic (Figures 17–18). Additionally, we added asset-ready comparisons (Figure 8, Table 6).

I appreciate the authors’ detailed and thoughtful responses, as well as the additional comparisons and supplementary results they provided. The explanations about the discrepancy between visual video quality and image metrics, and the distinction between raw meshes and post-processed meshes are reasonable.

That said, after reviewing the newly added figures, the overall improvement in rendering quality appears modest. Given this, I intend to keep my current score for this submission.

Reviewer_RxEGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PCAsplat, a geometric regularization for Gaussian splatting that produces surface-aligned Gaussian point clouds with far fewer floaters, so downstream geometry-processing tasks such as point-cloud segmentation and Poisson surface reconstruction can use the optimized Gaussian centers directly as input.

Specifically, building on the classical PCA-based normal estimation, the method performs PCA on each Gaussian's local neighborhood to obtain the covariance eigenvalues and a local normal estimate, namely the eigenvector associated with the smallest eigenvalue. Two losses are built on these statistics: L_PCA regularizes the eigenspectrum (penalizing the smallest eigenvalue to flatten the neighborhood onto its tangent plane, and penalizing the gap between the two tangent eigenvalues to promote isotropic coverage), while L_normal aligns each Gaussian’s shortest axis with the PCA normal. Together, they constrain both the positions and the local distribution of the Gaussians.

Qualitative and quantitative results show that the reconstructed Gaussian centers are well distributed on the underlying geometric surface, and that this brings clear gains to point-cloud segmentation and Poisson reconstruction performed directly on the Gaussian centers.

Critical strengths
  1. The motivation is clear and well argued.
  2. This method is simple, easy to follow, and can be easily added to existing Splatting pipelines.
  3. The experiments show a clear reduction in the error of the Gaussian centers, especially in the DTU dataset, and the qualitative results confirm that the Gaussians are well distributed on the underlying surface with fewer floaters.
  4. Directly using the optimized Gaussian centers as input point clouds yields clear gains in 3D instance segmentation on ScanNet++ and cleaner Poisson surface reconstructions.
Critical weaknesses
  1. The geometric accuracy of the Gaussian centers is still limited at scene scale. On Tanks and Temples, the F1-score of PCAsplat’s centers (0.387, Tab. 3) is higher than that of other methods’ centers, but it remains far below the F1-score of the meshes extracted by the same pipelines (0.50 for PCAsplat and 0.54 for GW, Tab. 7). This undermines the paper’s claim that the centers can be used directly for downstream tasks such as segmentation and surface reconstruction. More importantly, if a point cloud is all that a downstream task needs, the paper does not explain why one should use the Gaussian centers rather than the point cloud obtained by rendering and fusing the depth maps of the same Gaussian model (or by sampling its TSDF mesh). This is a free by-product of any splatting pipeline, and Tab. 7 indicates that it is more accurate than the centers at scene scale. This obvious alternative is missing from Tab. 3 and Tab. 5.

  2. The gains in surface reconstruction and novel view synthesis are limited.

    (a) Compared with the RaDe-GS (baseline PCAsplat builds on), PCAsplat improves DTU mesh CD only marginally (0.485 → 0.471 in Tab. 6) and brings no gain on Tanks and Temples (0.50 vs. 0.52 in Tab. 7), despite a ~54% longer training time. The footnote of Tab. 7 seems contradictory: the authors say they could not re-run RaDe-GS mesh extraction due to compute constraints, yet PCAsplat uses the same TSDF pipeline from the same codebase, and they did run RaDe-GS and got 0.49. The author should specify the extraction method and parameters.

    (b) On the outdoor scenes of Mip-NeRF 360 (Tab. 8), PCAsplat falls clearly behind RaDe-GS. This is not surprising. Gaussians that lie off the surface are not necessarily harmful: they still contribute view-dependent appearance in regions such as specular highlights and semi-transparent materials. Moreover, for fine structures such as vegetation, the notion of a surface itself is ill-defined, and I suspect the proposed regularization may actually hurt in these regions. The paper lacks any analysis or explanation of this, and only mentions the NVS results in passing.

  3. The baseline in Tab. 4 is unclear. The 0.655 in the first row does not match the 0.662 reported for RaDe-GS in Tab. 1 and Tab. Moreover, the improvement from L_normal is within the noise, and the contribution is not supported by Tab. 4.

  4. The ScanNet++ segmentation experiment is hard to interpret. The evaluation protocol is not described. The authors need to clarify which scenes are used, how the metrics are computed, and so on. I am also puzzled by the oracle result. Why does it reach only about 30%? Does this suggest that the choice of segmentation method itself is problematic?

  5. There are several typos and inconsistent names (e.g., “ADITIONAL”, “GaussianWrapping” vs. “GW”, “MILo” vs. “MiLo”).

Questions for authors

Overall, my main concern is the practical value of the method. For downstream tasks such as segmentation and Poisson reconstruction, a point cloud obtained from the rendered depth of the same model seems to be a better choice, and the proposed method does not show a clear advantage in this respect. In my opinion, the most valuable application of this work is the potential it shows for editing Gaussian assets, where surface-aligned splats can be cropped and reused without manual floater removal. However, this is only briefly demonstrated with qualitative results in Appendix C. I would encourage the authors to make this the central application of the paper and to support it with quantitative evidence, for example the amount of manual cleanup saved, the rendering quality of the cropped objects, or a comparison with existing Gaussian editing or segmentation pipelines. A convincing demonstration here would substantially strengthen the paper.

In addition, for volumetric 3D Gaussians such as those in GaussianWrapping or GGGS, the rendered iso-surface is biased toward the camera and the centers sit slightly behind the surface. Since the method is claimed to be plug-and-play, I would like the authors to discuss how it works with such methods.

3 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

Thank you for the constructive feedback. We have added the requested depth/mesh alternatives, quantitative asset evaluation, and protocol clarifications. We will also correct the typos and standardize method names.

(W1) “the paper does not explain why one should use the Gaussian centers rather than the point cloud obtained by rendering and fusing the depth maps of the same Gaussian model (or by sampling its TSDF mesh).”

Appendix E.1, Tab. 10 compares Gaussian centers, depth-derived points, and TSDF mesh samples on the same eight ScanNet++ scenes. Every input cloud contains 732k points and is evaluated using the same frozen Mask3D model. PCAsplat’s Gaussian centers achieve the highest Top-1 and Top-3 mIoU among the evaluated inputs.

InputPCA Top-1RaDe Top-1PCA Top-3RaDe Top-3
Gaussian centers (732k points, Tab. 5)24.357.7724.847.81
Depth points21.2220.4821.2220.50
Raw TSDF mesh samples21.5621.8721.2521.86
Post-processed mesh samples20.5923.0120.7422.78

These results support using the optimized centers directly, without depth rendering, backprojection, TSDF fusion, mesh extraction, or mesh post-processing, which can be time- and compute-heavy.

We also evaluated raw TSDF meshes before connected-component filtering, a post-processing step used to remove disconnected fragments caused by floaters (Fig. 17). On NeRF Synthetic, mean F1@5mm increases from 0.6199 to 0.6273, while the median number of floating mesh components outside the 5mm GT margin falls from 693.5 to 44.5 (93.6% fewer). Fig. 18 shows that this filtering can also remove valid geometry, including instruments in Drums and foliage in Ficus. Therefore, our resulting Gaussian representation provides surface samples directly for point-cloud applications.

(W2) “On T&T, the F1-score of PCAsplat’s centers (0.387, Tab. 3) is higher than that of other methods’ centers, but it remains far below the F1-score of the meshes extracted by the same pipelines (0.50 for PCAsplat and 0.54 for GW, Tab. 7).”

PCAsplat achieves the highest F1 among the evaluated Gaussian-center point clouds on T&T (0.387 vs. 0.321 for MILo). Fig. 13 illustrates the Gaussian-center distributions on these scenes. The gap to extracted meshes (0.50 for PCAsplat; 0.54 for GW) can reflect differences in sampling density and coverage (F1’s recall and precision): Gaussian centers may leave parts of a surface sparsely sampled, particularly where large Gaussians regions, even when their centers are close to the surface. The segmentation comparisons in Tab. 10, using 732k points per cloud and the same frozen Mask3D model, further support the practical value of Gaussian centers for downstream processing without depth rendering or mesh extraction.

(W3) “PCAsplat improves DTU mesh CD only marginally (0.485 → 0.471 in Tab. 6) and brings no gain on T&T (0.50 vs. 0.52 in Tab. 7), despite a ~54% longer training time.”

TSDF meshes are fused from rendered depth, where occluded or weakly contributing floaters have little influence. Thus, reducing these floaters can improve Gaussian geometry without substantially changing mesh scores. We evaluate these directly using Gaussian-center accuracy and downstream tasks (Tabs. 1–3, 5, 6, and 10; Figs. 8–9), with the supp. videos also showing fewer visible floaters.

These come with a trade-off in training time: on DTU, ours takes 4.98 minutes versus 3.38 for RaDe-GS on the RTX PRO 4500 Blackwell, and our reported time is 22.26 minutes versus 10.75 min on the RTX A5500. Further optimization of our implementation could reduce this overhead. Additional details are given in Appendix E.4, Tab. 14, and our response to A3ZJ.

(W4) “The author should specify the extraction method and parameters.” (Tab. 7)

We reran both methods using TSDF fusion with identical, GT-independent settings, obtaining 0.50 F1 for PCAsplat and 0.49 for our reproduced RaDe-GS (Tab. 8). We used 0.002 voxel size and a 5 max depth distance and kept only the largest component. RaDe-GS's published 0.52 uses GT-derived bounds from PGSR's transforms.json, which restrict the extraction volume and enable finer resolution at lower cost. Reproducing that resolution without GT-derived bounds was computationally prohibitive.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

(W5) “On the outdoor scenes of Mip-NeRF 360 (Tab. 8), PCAsplat falls clearly behind RaDe-GS.”

In Fig. 16 we compare PCAsplat and RaDe-GS on two indoor and two outdoor scenes, and observe very similar foliage detail and reflective appearance.
We couldn’t reproduce RaDe-GS’s published Mip-NeRF 360 results using its official training script. Tab. 9 reports the published results, our reproduced baseline, and PCAsplat:

MethodIndoor PSNR / SSIM / LPIPSOutdoor PSNR / SSIM / LPIPS
RaDe-GS, published30.74 / .928 / .16525.17 / .764 / .199
RaDe-GS, reproduced30.77 / .930 / .17024.84 / .740 / .240
PCAsplat30.46 / .930 / .17024.70 / .730 / .250

Against the reproduced baseline, the outdoor PSNR gap is 0.14 dB (24.84 vs. 24.70), rather than 0.47 dB, with PCAsplat’s SSIM lower by 0.010 and LPIPS higher by 0.010. We also added a visual comparison in Fig. 16 between RaDe and PCAsplat on Mip-NeRF 360 .

(W6) “The baseline in Tab. 4 is unclear. The 0.655 in the first row does not match the 0.662 reported for RaDe-GS in Tab. 1 and Tab. Moreover, the improvement from L_normal is within the noise, and the contribution is not supported by Tab. 4.”

Tab. 1 reports the original RaDe-GS implementation (0.662), while Tab. 4 uses our modified baseline (0.655) with our NCC weight 0.6 instead of 0.3. PCA reduces CD from 0.655 to 0.520; the full model reaches 0.519. The normal term aligns each Gaussian's smallest axis with the PCA normal and encourages thin primitives (Eq. 4), with more coherent normal directions shown in the DTU apple slice (Fig. 19). Gaussian-center CD does not measure normal orientation, which we use alongside the centers for direct Poisson reconstruction (Fig. 9).

(W7) “The evaluation protocol is not described. The authors need to clarify which scenes are used, how the metrics are computed, and so on. I am also puzzled by the oracle result. Why does it reach only about 30%?”

Appendix A.1 describes the protocol and lists the eight scene IDs. A frozen Mask3D pretrained on ScanNet200 predicts instance masks, class labels, and confidence scores. We map classes to a common label set, transfer GT labels from nearest GT vertices to the reconstructed points (Gaussian centers), and compute per-class IoU identically for all methods.

The same model achieves 30.41% Top-1 mIoU on GT geometry, providing a reference performance level rather than a perfect-label oracle or theoretical upper bound. PCAsplat achieves 24.35%, versus 10.54% for the strongest original reconstructed-cloud baseline (Tab. 5).

(Q1) “support it with quantitative evidence, for example the amount of manual cleanup saved, the rendering quality of the cropped objects”

We have added the asset evaluation to Sec. 4.3 (Tab. 6, Fig. 8). On Ignatius, Caterpillar, Barn, and Truck, cropped Gaussians are rendered from test viewpoints covering the entire object and compared against images masked using projected GT geometry. Predictions remain unmasked, so unwanted content outside the asset is penalized.

MethodPSNRSSIMLPIPS-VGG
RaDe-GS24.418.9467.0582
GaussianWrapping23.188.9238.0733
MILo22.797.9293.0686
PCAsplat25.021.9495.0571

PCAsplat performs best on all three metrics. In a timed comparison, removal took approximately 90 seconds for RaDe-GS versus 10 seconds for PCAsplat. We note that low-opacity floaters are less invisible in photo renders; to evaluate this case, Figs. 11–13 reveal them as Gaussian centers. The extracted Gaussian assets can be reused directly without geometry conversion.

(Q2) “Since the method is claimed to be plug-and-play, I would like the authors to discuss how it works with such methods.” (GaussianWrapping and GGGS)

Our PCA loss regularizes relative neighborhood geometry and is invariant to common translations; it imposes no hard constraint that centers lie on the rendered isosurface, as this is defined by the renderer. GGGS renders median depth, while GaussianWrapping and MILo support pivot-based marching tetrahedra. Tab. 2 demonstrates integration into 2DGS, PGSR, MILo, and RaDe-GS without changing their rasterizers. The full objective, including normal alignment and scale penalties, remains to be validated with GGGS and GaussianWrapping.

Thanks for the detailed response and the additional experiments.

The Gaussians are clearly cleaner, but the downstream gains over RaDe-GS remain modest. On segmentation (Tab. 10), PCAsplat centers reach 24.35% Top-1 mIoU, versus 23.01% for RaDe-GS mesh samples, which come for free with the baseline. On asset rendering (Tab. 6), PSNR improves from 24.42 to 25.02 dB, with only small changes in SSIM and LPIPS. Given that training takes about 1.5–2× longer (Tab. 14), the practical benefit over the baseline seems limited.

The responses address several of my concerns, and I will take them into account in the next discussion phase. Overall, I see the paper as borderline and would not object to acceptance.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. The revised text is highlighted in blue.

We have added Additional experiments:

  • Expanded cropped-asset evaluation: Tab. 6, Fig. 8.
  • Mip-NeRF 360 Gaussian-center comparisons: Fig. 11.
  • PGSR and RaDe-GS mesh comparisons: Fig. 15.
  • Reproduced RaDe-GS rendering metrics and comparisons: Tab. 9, Fig. 16.
  • Raw meshes and post-processing effects: Figs. 17–18.
  • Gaussian centers vs. depth and mesh inputs: Tab. 10.
  • Loss ablations and centroid smoothing: Tabs. 11–12.
  • Normal-alignment comparison: Fig. 19.
  • Neighborhood sensitivity: Tab. 13.
  • Same-GPU training times: Tab. 14.

Added further clarifications

  • Ablation baseline and NCC weight: Sec. 4.2.
  • Segmentation protocol: Appendix A.1.
  • Controlled TSDF extraction settings: Appendix C.1.

And strengthened our discussion:

  • Gaussian geometry vs. mesh and rendering quality: Appendix C.
  • Renderer independence of PCA regularization: Appendix D.
  • Direct use of Gaussian centers: Appendix E.1.
  • Neighborhood-size trade-offs: Appendix E.3.

We thank the reviewers again and welcome further discussion.

AuthorsOpenReview ↗

Updated manuscript

  • Updated Fig. 19 for reviewer A3ZJ response

Dear reviewers,

The authors have submitted responses to your reviews, yet no feedback has been provided. Could you please give your feedback and engage with the authors if there is any unclearness?

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.