acceptodds
Under review as a conference paper at ICLR 2027

Group-Wise Simplification of Radiative Gaussian Splatting for Computed Tomography Guided by Convex Density Optimization

Abstract

Radiative Gaussian splatting achieves high fidelity computed tomography (CT) reconstruction, but heuristic densification causes severe primitive redundancy. Pruning redundant primitives is challenging because naive removal damages subtle structural boundaries and induces streak artifacts. To resolve this issue, we propose CDGS, a structure preserving group-wise simplification framework guided by convex density optimization. By exploiting the linearity of CT projection physics under fixed geometry, we formulate candidate pricing as a convex quadratic subproblem solved via sketching. Primitives are partitioned into topological structural groups that confine density compensation within contiguous structural boundaries, which prevents non local ray interference. In addition, an autonomous in-group stopping rule dynamically terminates simplification when redundancy is exhausted. Experiments across multiple CT benchmarks demonstrate that CDGS substantially reduces primitive counts while maintaining reconstruction fidelity. Our framework outperforms existing state-of-the-art methods in both numerical accuracy and structural preservation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, CT Reconstruction, Gaussian Primitive Simplification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7bk6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes CDGS, which removes the extra Gaussians group by group so that only nearby ones adjust to fill each gap, which avoids blurred edges and streak artefacts. It cuts the number of Gaussians by nearly half with almost no loss in image quality, very slightly beating other pruning methods.

Critical strengths

The paper demonstrates a well-motivated CT-specific formulation, valid convex subproblem analysis, comparisons on two backbones, and additional fixed-count experiments in the appendix. Note that Appendix is quite long.

Critical weaknesses
  • Section 4.4.2, how to ensure that the solver converges.
  • CDGS re-fits densities with NNLS after every batch and its three events are interleaved with further training. The paper never says when the baselines are applied, or whether they get the same density re-fit or fine-tuning afterwards. So, comparisons may be unfair.
  • The visual differences between methods in Figure 4 are difficult to distinguish, even in the enlarged regions. The figure therefore does not clearly demonstrate the claimed improvements in fine-structure preservation and artifact suppression. Please provide larger, higher-resolution crops with consistent intensity windows, mark the specific structures or artifacts being compared, and include error maps or intensity profiles to make the differences easier to assess.
  • Table 2 suggests grouping provides a slight benefit, while convex-cost selection contributes more. Without convex-cost selection, grouping changes PSNR by only about approx. 0.01 dB. Its benefit therefore appears mainly when combined with convex-cost selection. This seems that the specific benefit of grouping for boundary preservation and streak suppression is unclear. Please provide targeted structural measurements and comparisons with simpler spatial grouping methods to better justify this central component.
  • The paper provides some local complexity discussion and runtime measurements for different sketch dimensions, but the overall computational cost remains unclear. Please report total runtime and peak memory relative to the backbone and competing methods, and explain how the cost scales with the number of Gaussians, projection views, and simplification rounds.
Questions for authors

Please see weaknesses above.

10 comments
AuthorsOpenReview ↗

Response to Reviewer 7bk6 (1/4): W1

We thank the reviewer for the careful reading and the constructive comments. The changes in the revised PDF are listed in our general comment “Summary of the revision”. All section, equation, figure, and table numbers below refer to the revised PDF, in which new and revised text is highlighted in blue. Figures cannot be posted here, so we point to the revised PDF for them and reproduce the key tables.

Response to W1 (convergence of the solver, Section 4.4.2)

Thank you for raising this. In short, the pricing problem is convex, so the iteration in Eq. (10) converges to its global minimum for every candidate. This comes from the structure of the problem, not from tuning. Section 4.4.2 contains the ingredients but did not spell out the argument. We have added it (Appendix B.5).

The problem has no bad local minima. The solver adjusts densities only. Positions and shapes are held fixed during the solve. Under fixed geometry the projection is linear in density, so the objective in Eq. (8) is

qa(α)=12P∥s(a)+Saα∥2.q_a(\alpha) = \frac{1}{2P}\lVert s(a) + S_a\alpha \rVert^2 .

This is the square of a function that is linear in α\alpha. Along any straight line it is a parabola that opens upward or is flat. Such a function is convex, and the lower bounds in Eq. (9) keep the feasible set convex. In a convex problem, a point where no further descent is possible is a global minimum. This is the difference from optimizing positions and shapes, which is non-convex and stays with the backbone training.

The iteration cannot overshoot. We use projected gradient descent, Eq. (10). On a convex quadratic it can only fail if the step size is too large. The safe limit is 1/L1/L, where LL is the largest eigenvalue of the Hessian GaG_a. We set τ=1/tr(Ga)\tau = 1/\mathrm{tr}(G_a). The eigenvalues of GaG_a are non-negative, so tr(Ga)≥L\mathrm{tr}(G_a) \ge L and τ≤1/L\tau \le 1/L for every candidate (Appendix A.4.1).

Guarantee. With these two facts, the standard result for projected gradient descent (e.g., Beck, 2017, Ch. 10) applies:

qa(α(t))−qa⋆≤tr(Ga)∥α(0)−α⋆∥22t.q_a(\alpha^{(t)}) - q_a^\star \le \frac{\mathrm{tr}(G_a)\lVert \alpha^{(0)} - \alpha^\star \rVert^2}{2t} .

No iteration increases the objective, and the gap to the global minimum goes to zero.

Uniqueness. GaG_a is positive semidefinite, not positive definite, so the minimizer may not be unique. The minimal value is unique, and the price c(a)c(a) depends only on that value.

In practice. We run Tq=20T_q=20 iterations from α(0)=0\alpha^{(0)}=0 (Appendix A.4.1). Every iterate is feasible and the objective never increases, so the returned price is at or above the exact minimum, never below.

Table A13 (reproduced below) checks whether 20 iterations are enough. (a) Compared with converged prices, the 20-step prices keep the ordering that the selection relies on: the rank correlation is 0.83–0.99, and 83–89% of the cheapest 10% of candidates coincide, although a median gap of 0.20–0.31 remains in the values. (b) The final result is stable from 20 iterations on: PSNR is 37.95, 37.95 and 37.97 dB for Tq=20T_q=20, 50 and 100. Fewer iterations inflate the prices, and with them the acceptance tolerance of Eq. (12), so more primitives are removed at lower quality (37.52 dB at Tq=1T_q=1). More iterations make the stop more conservative (33.14k and 34.37k primitives) for at most 0.02 dB, while pricing time stays at 44–52 s.

Table A13(a). 20-step prices against converged prices on both backbones (50 views). The gap is (q20−q⋆)/(q0−q⋆)(q_{20}-q^\star)/(q_0-q^\star), where q0q_0 is the price before any compensation and q⋆q^\star the optimal price, so 0 means converged and 1 means no progress.

BackboneActionRemaining gap (median)Rank corr.Cheapest-10% overlap
FaCT-GSprune0.1970.8590.837
FaCT-GSmerge0.2920.9840.892
R²-Gaussianprune0.2190.8260.832
R²-Gaussianmerge0.3060.9850.882

Table A13(b). Sweep of TqT_q on FaCT-GS with 50 views; #G in thousands; the paper uses Tq=20T_q=20.

TqT_qPricing roundsPricing (s)Total (s)#G (k)PSNR
18643.8435.128.4137.52
109147.0434.028.6637.75
207546.1424.930.2437.95
508852.4469.233.1437.95
1008951.7454.434.3737.97

Batch re-fit. Eq. (A4) solves the same kind of convex problem with an accelerated iteration whose step size uses an estimated largest eigenvalue. We do not rely on monotone descent there: the solver returns the best feasible iterate, and the batch is accepted only if its objective passes Eq. (12). A poorly converged re-fit can cause a rejection, but not an accepted degradation.

AuthorsOpenReview ↗

Response to Reviewer 7bk6 (2/4): W2, W3

Response to W2 (fairness of the comparison)

Thank you for checking this. Section 5.1 gives the reconstruction budgets and the CDGS schedule, and we clarify their role below. The protocol is now also stated in Appendix B.6.

Same number of optimization steps. All methods are trained on the same backbone for the same total number of steps: 30,000 for R²-Gaussian and 20,000 for FaCT-GS. The simplification events of CDGS do not consume training steps, and we add no extra fine-tuning steps after them. CDGS therefore uses the same number of backbone optimization steps as the corresponding baseline.

Schedule of CDGS. All three events take place after densification has ended (steps 16,001, 19,001 and 22,001 for R²-Gaussian). No new primitives are created after the events, so the reported count is the final count.

Schedule of the baselines. NanoGS, LightGaussian, GHAP, and PUP 3D-GS take a target primitive count. We apply these four methods at the same three iterations as CDGS: 11,001, 14,001, and 17,001 for FaCT-GS, and 16,001, 19,001, and 22,001 for R²-Gaussian. At each event, we use the count retained by CDGS as the target for each baseline. All five methods then resume backbone training for the same number of steps. CDGS additionally performs candidate pricing and NNLS density re-fitting; the baselines do not receive the CDGS re-fit. The comparison therefore matches the schedule, count targets, and subsequent backbone training, but not the total computation. Under this schedule, CDGS is 0.09 to 1.01 dB higher than these four baselines (Table 1 and Table A6). GR-Gaussian, LeGS, FastGS, and Perceptual-GS are not post-hoc pruning methods. They act through the density control during training, so there is no application step and no fine-tuning stage to schedule. They are trained for the same total number of steps and evaluated at the final step. We do not force a primitive count on them, because that would change the methods. Their final counts are reported in Table 1 and Table A6.

Density re-fit. The baselines do not receive the NNLS re-fit. It is not a post-processing step granted to CDGS alone. It is part of the proposed method. It follows from the central observation of this paper: under fixed geometry, the CT projection is linear in density, so the densities can be re-solved as a convex problem. Using this property for simplification is our contribution. Existing simplification methods were designed for nonlinear alpha blending and have no counterpart to this step.

In Table 2, the “w/o Convex Cost” variant uses the same events and the same primitive count at each event as CDGS. It only replaces cost-based selection with random selection. It is 0.26 to 0.60 dB below CDGS. This shows that the convex-cost selection itself contributes clearly.

Response to W3 (local details and reconstruction artifacts, Figure 4)

Thank you for pointing this out. We agree that the original Figure 4 was too small, and the differences between methods were hard to see. The figure did not serve its purpose.

Part of the reason is real. At matched primitive counts, the methods are close in overall quality. In the three cases of Figure 4, all methods except LeGS are within 1 dB of each other, so whole slices look alike. The differences sit in local fine structures and at boundaries, and the original figure did not enlarge them enough.

Following your suggestion, we show larger crops for Head and Present in Figure 5 and for all 18 objects in Figure A10 (Appendix B.3), on FaCT-GS with 25 views. Both figures also include CDGS without grouping (“w/o Grp.”). Cyan boxes locate the crops; green arrows mark the same location across methods. For each object, the reconstructions share an intensity window, and the absolute-error maps share a color scale. The numbers are whole-volume PSNR values (dB), not crop PSNR.

In Head, the arrow marks the uniform region next to the bone; in Present, it marks the flat region below the ring. The streaks visible in several baseline reconstructions are less pronounced in the corresponding CDGS crops.

The quantitative results agree with these local differences. At matched primitive counts, CDGS has the highest PSNR and SSIM in all three view settings (Table 1). SSIM measures structural similarity.

AuthorsOpenReview ↗

Response to Reviewer 7bk6 (3/4): W4

Response to W4 (role of grouping, Table 2)

Thank you for this comment. The pattern in Table 2 is real, and our original text did not explain its cause.

Grouping and convex cost form one mechanism. CDGS is not a baseline with two independent modules added on top. Grouping sets the scope in which candidates are priced and in which densities may compensate. Convex cost is what is computed inside that scope. In the “w/o Convex Cost” row, selection is random and nothing is priced, so grouping has nothing to constrain. Random selection inside groups and random selection over the whole volume are almost the same procedure. This is why the two rows differ by only 0.01 dB. It follows from the design and does not mean that grouping has no effect.

The synergy shows in the other direction of Table 2. With convex cost, grouping adds 0.15 / 0.10 / 0.07 dB at 75 / 50 / 25 views. The gain from convex cost is 0.60 / 0.47 / 0.26 dB with grouping and 0.46 / 0.38 / 0.18 dB without it. Grouping lets the convex cost work more effectively. Section 5.3 now states this explanation.

We also verified directly whether the compensation stays inside anatomical structures. Figure 6 (Figure 5 of the original submission) gives the qualitative result. Without grouping, the residual of a pruned primitive is absorbed by distant primitives along intersecting rays. With grouping, the compensation is confined to the same structure. We have added a quantitative experiment that measures the fraction falling outside the corresponding structure (out-of-structure rate). The results are given in Table A9 (Appendix B.1) and Figure 7, with the complete comparison in Figure A8. Our grouping has a clearly lower rate than the simpler spatial grouping methods: k-means, KD-tree, agglomerative clustering, and DBSCAN. In addition, Figures 5 and A10 now include CDGS without grouping (“w/o Grp.”) with error maps: on FaCT-GS with 25 views, grouping gives the higher PSNR on 15 of the 18 objects, ties on one, and is lower on two by at most 0.05 dB.

Reviewer_j37rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper formulates density fitting under fixed Gaussian geometry as a convex optimization problem, combined with structural grouping and adaptive stopping for Gaussian simplification is novelty. However, the detailed design of group-wise is a bit confusing.

Critical strengths
  1. Exploiting the linear projection model to formulate density fitting as a convex non-negative problem.

  2. The method reduces the number of Gaussians by nearly 50% while maintaining competitive PSNR.

Critical weaknesses
  1. Group-wise approach has a reasonable geometric motivation. However, the choice of Bhattacharyya distance, Jaccard similarity, and Leiden clustering is not sufficiently justified. Geometric similarity does not necessarily correspond to anatomical structural coherence, particularly for Gaussians with substantially different scales. Comparisons with simpler grouping strategies would help establish the necessity of this design.

  2. While the convex density optimization exploits CT projection physics, the proposed structural grouping relies primarily on geometric similarity based on Gaussian positions and covariances.

Questions for authors

1.Why is Bhattacharyya distance preferred over simpler spatial or Gaussian-overlap-based similarity measures?

2.How does the grouping behave when large Gaussians overlap multiple smaller primitives or structural boundaries?

  1. Since structural grouping relies on Gaussian geometry and Leiden community detection rather than anatomical labels, could the authors clarify to what extent graph connectivity corresponds to anatomical boundaries? Could quantitative boundary results be provided?

4.The proposed grouping restricts density compensation during simplification, while standard reconstruction training resumes afterward. Could the authors clarify whether the structural benefits of group-wise simplification are preserved after subsequent optimization?

3 comments
AuthorsOpenReview ↗

Response to Reviewer j37r (1/3): W1

We thank the reviewer for the careful reading and the constructive comments. The changes in the revised PDF are listed in our general comment “Summary of the revision”. All section, equation, figure, and table numbers below refer to the revised PDF, in which new and revised text is highlighted in blue. Figures cannot be posted here, so we point to the revised PDF for them and reproduce the key tables.

Response to W1

Thank you for recognizing the value of the convex density formulation and for asking how closely our geometric groups follow structural boundaries. Your questions prompted us to compare simpler grouping methods and to measure agreement with intensity-defined reference regions. We agree that the paper did not justify these choices well enough.

Grouping is not a segmentation step. It is a constraint on the convex criterion. The convex density cost measures only projection error. Under sparse views, a distant primitive on the same rays can cancel the residual left by a removed primitive. The cost then looks small although the local structure has changed. Grouping decides which primitives may adjust their densities when an action is priced. The price then reflects the quality loss that local neighbors cannot absorb.

This role sets two requirements. First, the number of groups cannot be preset. We do not know how many structures an object contains. The number varies across objects and changes during simplification, because groups are rebuilt after each accepted round. Second, each group must be connected. A group with two separate parts would allow compensation across the gap, which is exactly what grouping should prevent.

The design follows from these requirements. A kNN graph with Jaccard weighting and community detection is a standard pipeline for clustering points with uneven density and an unknown number of clusters, e.g., PhenoGraph (Levine et al., 2015) and Seurat (Stuart et al., 2019). The Jaccard term down-weights edges between primitives that share no neighbors. Such edges are usually bridges between two structures. Leiden needs no preset group number and guarantees connected communities, which Louvain does not (Traag et al., 2019). The group-wise simplification framework is not tied to a particular pairwise distance. Leiden operates on the resulting weighted graph without requiring a prescribed number of groups, and its resolution parameter controls the grouping granularity. The result is insensitive to this parameter: for γ∈{1,2,4,8,16}\gamma \in \lbrace 1,2,4,8,16 \rbrace, GEP stays within 0.538–0.547 (Table A2; FaCT-GS, 50 views).

Table A9 (reproduced below; per-view results in Table A8) compares our grouping with simpler strategies: Euclidean distance, Mahalanobis distance, k-means, KD-tree, agglomerative clustering, and DBSCAN. CDGS maintains comparable PSNR to Euclidean and Mahalanobis grouping while using fewer Gaussians and achieving lower mean GEP on both backbones. DBSCAN reaches a lower GEP only by removing far more primitives, at a cost of 1.14 dB on FaCT-GS and 0.63 dB on R²-Gaussian. Grouping takes 12.6 s on average, about 2.09% of the average training time. In the revision, we moved the formulas to the appendix (Appendix B.7) and kept the motivation in the main text.

Table A9. Averages over 25/50/75 views at the end of training; #G in thousands. The out-of-structure rate (lower is better; protocol in Q3) is averaged over all 18 objects. “No grouping” is the w/o Grouping ablation at the primitive count of CDGS. Euclidean and Mahalanobis replace only the distance; k-means, KD-tree, Agglo., and DBSCAN replace the grouping algorithm. Grouping time is averaged over views and both backbones.

MethodFaCT-GS PSNR#G (k)GEPR²-Gaussian PSNR#G (k)GEPOut of structure (%)Grouping time (s)
No grouping37.1130.420.58037.4135.770.549––
Euclidean37.3332.440.58837.5138.670.58029.811.7
Mahalanobis37.3432.390.58637.5036.870.55525.911.8
k-means37.3532.030.57837.5639.140.58135.736.9
KD-tree37.3732.900.59137.5740.260.59732.837.0
Agglo.37.3431.020.56137.5037.880.57035.786.5
DBSCAN36.2020.690.49436.8827.560.47947.3132.8
CDGS (Ours)37.3430.420.55037.5135.770.53515.512.6
AuthorsOpenReview ↗

Response to Reviewer j37r (2/3): W2, Q1, Q2

Response to W2

We agree that the grouping uses only positions and covariances, and this is deliberate. The convex cost is where the CT physics enters, but under sparse views it cannot tell apart primitives that are far apart yet change the projections alike. A grouping built from projection responses would inherit this ambiguity, so it has to come from information the projections do not provide, namely spatial geometry. Inside a group the physics is used again: the adjustable set is chosen by the correlation of sketched projection responses (Eq. (A11)). Although no intensity is used, 84.5% of the group members fall inside one intensity-defined structure (Q3). Q1–Q3 give the details.

Response to Q1 (why Bhattacharyya distance)

Thank you for this comment. We chose Bhattacharyya distance precisely because it measures Gaussian overlap without requiring a cutoff for either kernel. For normalized spatial kernels gig_i and gjg_j, the overlap coefficient is

exp⁡(−DB)=∫gi(x)gj(x)dx∈[0,1].\exp(-D_B)=\int\sqrt{g_i(\mathbf{x})g_j(\mathbf{x})} \mathrm{d}\mathbf{x} \in[0,1].

Its negative logarithm is the distance in Eq. (A32) (Appendix B.7). Euclidean center distance ignores scale and orientation. The first term of Eq. (A32) is one eighth of the squared Mahalanobis distance under the mean covariance; it is zero for concentric kernels regardless of their sizes. The covariance term distinguishes such kernels. For example, two concentric isotropic 3D Gaussians with a 10:1 standard-deviation ratio have zero center distance but an overlap coefficient of only (20/101)3/2≈0.09(20/101)^{3/2}\approx 0.09. Unlike the KL divergence, it is symmetric, so it directly defines an undirected edge weight.

The Euclidean and Mahalanobis rows of Table A9 change only the distance, keeping graph construction and candidate pricing fixed (averaged over 25/50/75 views).

The distance has little effect on PSNR. It changes how well the groups follow the structures and how many Gaussians can be removed. Bhattacharyya gives the lowest out-of-structure rate on all 18 objects (see Q3). It also gives the fewest Gaussians and the lowest GEP on both backbones at comparable PSNR. Per setting, it has the lowest GEP in five of six, and Mahalanobis is lower in one (R²-Gaussian, 25 views). The extra grouping time is about 1 s.

Response to Q2 (large Gaussians over small primitives or boundaries)

Geometric overlap alone does not determine group membership.

First, both backbones split large, high-gradient Gaussians during densification, but some surviving Gaussians still span a boundary.

Second, for each object and view count, we define the largest 5% of Gaussians as the large subset. Both bars in Figure 8 count members of this subset that cross an intensity-defined boundary. Dark blue reports their percentage within the large subset; light blue reports their percentage among all Gaussians. Across the eight settings, the latter ranges from 0.5% to 4.5%. For Jaw at 50 views, about 90% of the large subset crosses a boundary, corresponding to 4.5% of all Gaussians.

Third, Figure 8 (left; Chest, Foot, and Head at 50 views, R²-Gaussian backbone) shows what happens in these examples. Yellow boxes locate the enlarged regions, white contours indicate reference structural boundaries, and colors denote groups. The thick blue outlines represent the footprints of the illustrated large Gaussians.

Although the large Gaussian geometrically overlaps primitives across the boundary, the small primitives on the other side retain their own groups. The illustrated large Gaussian therefore does not bridge the two structures into one group. The Bhattacharyya distance penalizes the scale difference, while mutual neighbors with Jaccard weighting suppress weak links. Its removal is then priced with in-group compensation only and must pass the stopping and batch-acceptance criteria. A large Gaussian can also compensate for a pruned neighbor in its own group. Its density change alters the projections over its whole footprint, including the part across the boundary. This is part of the residual in Eq. (8) and can only be absorbed inside the group, so such leakage raises the price instead of hiding it.

More examples are shown in Figure A9 (Appendix B.2).

AuthorsOpenReview ↗

Response to Reviewer j37r (3/3): Q3, Q4

Response to Q3 (correspondence to boundaries)

Our grouping deviates least from the structure boundaries among the tested alternatives in the evaluated regions. We added a quantitative analysis across all 18 objects (Figure 7 for three objects, Figure A8 for all) and the out-of-structure column of Table A9.

Protocol. We segment each reference volume into structures by intensity thresholding. Each group is assigned to the structure that contains most of its Gaussians. We report the out-of-structure rate among the displayed primitives in the evaluated regions: the percentage whose reference-region label differs from the most frequent reference-region label in their group. Lower is better. The evaluated region of each object is the same for all methods. Reference regions are used only for evaluation, not to construct the groups. To control grouping granularity, we used the number of groups produced by Ours on each checkpoint as the target. We matched this number for k-means, agglomerative clustering, and KD-tree partitioning. For DBSCAN, we fixed the remaining parameters and selected the neighborhood radius to obtain the closest group count, without using reference-region labels.

Our grouping has the lowest rate on all 18 objects. On average, 84.5% of the Gaussians stay inside the structure of their group. The rate is 15.5%, against 25.9% for the best alternative and 32.8% to 47.3% for the other grouping algorithms.

Eucl. and Mahal. keep our pipeline and only replace the Bhattacharyya distance. Both raise the rate clearly. The other four replace the grouping algorithm, and all exceed 32%. This supports both design choices: the distance and the graph-based grouping.

The test is also strict. It uses only the center, and a kernel centered just outside a boundary still overlaps the structure.

In Figures 7 and A8, yellow boxes locate the crops, white contours mark the reference regions, and point colors identify groups within each panel. Red shading marks disagreement: a Gaussian center lies in a reference region other than its group's most frequent region. The percentage in each panel is the fraction of displayed centers with this mismatch. Figure A8 shows all 18 objects.

Response to Q4 (benefits after resumed training)

All metrics in the paper are measured at the end of training, after optimization resumes. The “No grouping” row of Table A9 therefore already compares grouped and ungrouped simplification after the subsequent optimization. On FaCT-GS, CDGS is 0.16 to 0.28 dB higher than the variant without grouping at the same Gaussian counts. GEP drops from 0.595/0.582/0.562 to 0.563/0.545/0.542 at 75/50/25 views. On R²-Gaussian, CDGS is 0.07 to 0.15 dB higher at the same Gaussian counts. GEP drops from 0.560/0.549/0.538 to 0.540/0.536/0.529.

Figures 5 and A10 now also show this variant as the column “w/o Grp.” next to ours, with error maps (FaCT-GS, 25 views, final checkpoint). CDGS has the higher PSNR on 15 of the 18 objects, ties on one, and is lower on two by at most 0.05 dB; the largest gains are 0.49 dB (Mount) and 0.43 dB (Teapot).

The reason is that every simplification round runs after densification has ended. No new primitives are created afterwards. The resumed training only refines the parameters of the surviving primitives. It cannot restore a removed primitive or undo a merge. Grouping decides which actions are taken. An action is cheap only if primitives in the same group can absorb it. Without grouping, actions that are absorbed only by distant primitives are also taken, and locally needed primitives are lost. This difference is fixed in the primitive set, so it remains at the end of training. The boundary results in Q3 are also measured at the end of training.

At matched primitive counts, grouped simplification retains a measurable PSNR advantage at the final checkpoint on both backbones. This shows that resumed training does not erase the observed quality advantage.

Reviewer_ofy5OpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SAGA for ultra-sparse-view CT reconstruction. It combines geometry-frequency regularization, structure-aware Gaussian allocation using volumetric gradients and residual backprojection, and a two-stage optimization schedule.

Critical strengths

The method is clearly motivated and integrates volumetric and projection-domain information in an intuitive manner.

SAGA consistently outperforms several analytical, neural-field, and Gaussian baselines under 6-, 8-, and 10-view settings.

Critical weaknesses
  1. Most improvement comes from conventional regularization. In Table 2, the objective raises PSNR from 26.06 to 30.76 dB, whereas G+P allocation adds only 0.13 dB. Under the same objective, G+P improves over gradient densification by merely 0.04 dB, making the central allocation contribution appear marginal.
  2. Evaluation uses only five simulated volumes, without realistic measurement noise, scanner variation, or real projection data. Moreover, reporting the checkpoint with the “best validation reconstruction quality” is unclear and may require reference volumes or additional unseen projections unavailable in the claimed ultra-sparse setting.
Questions for authors

Does checkpoint selection use reference volumes or additional held-out projections? How would the stopping criterion work with only the acquired sparse views?

Can the allocation comparison be repeated across all anatomies and view settings with multiple seeds and uncertainty estimates?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank all reviewers for their careful reading and constructive comments. We have uploaded a revised PDF. New and revised text is highlighted in blue. All section, equation, figure, and table numbers in our responses refer to the revised PDF. Figures cannot be posted here, so we point to the revised PDF for them and reproduce the key tables in the responses.

Main changes

  • Section 4.3. The three grouping formulas (original Eqs. (5)–(7)) are moved to Appendix B.7 (now Eqs. (A32)–(A34)), and the main text now states the two requirements behind the design. Later main-text equations therefore shift by three (e.g., original Eq. (11) is now Eq. (8)).
  • Section 5. Three figures are added: Figure 5 (enlarged local crops with error maps), Figure 7 (grouping strategies against structural boundaries), and Figure 8 (large Gaussians across boundaries). Section 5.3 now explains the pattern in Table 2. The original Figure 5 is now Figure 6; Figures 1–4 and Tables 1–2 keep their numbers.
  • New Appendix B (Additional Experiments and Analysis).
    • B.1 Comparison of grouping strategies: Tables A8–A9, Figure A8
    • B.2 Large Gaussians across structural boundaries: Figure A9
    • B.3 Full qualitative comparison with error maps for all 18 objects: Figure A10
    • B.4 Computational cost and scaling: Tables A10–A12
    • B.5 Convergence of candidate pricing: Table A13
    • B.6 Protocol of the compared methods
    • B.7 Formulation of structural grouping (the formulas moved from Section 4.3)
  • Appendix A is unchanged except for minor corrections.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.