- 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
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.
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.
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)=2P1∥s(a)+Saα∥2.
This is the square of a function that is linear in α. 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/L, where L is the largest eigenvalue of the Hessian Ga. We set τ=1/tr(Ga). The eigenvalues of Ga are non-negative, so tr(Ga)≥L and τ≤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⋆≤2ttr(Ga)∥α(0)−α⋆∥2.
No iteration increases the objective, and the gap to the global minimum goes to zero.
Uniqueness. Ga is positive semidefinite, not positive definite, so the minimizer may not be unique. The minimal value is unique, and the price c(a) depends only on that value.
In practice. We run Tq=20 iterations from α(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=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=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⋆), where q0 is the price before any compensation and q⋆ the optimal price, so 0 means converged and 1 means no progress.
Table A13(b). Sweep of Tq on FaCT-GS with 50 views; #G in thousands; the paper uses Tq=20.
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.