- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies novel-view synthesis for 3D Gaussian Splatting under a fixed storage budget. The core claim is that models with the same final Gaussian count need not follow the same growth trajectory — how the representation grows before pruning also affects the quality of the compressed model. The authors propose CAED, which combines an ImprovedGS-style frontend (edge-guided selection, long-axis splitting, and a square-root growth budget) with the PWRS compression backend (reconstruction-aware RPS/SPS pruning and DoG refinement), configuring the training-time growth budget B(t) separately from the final target capacity N_tar. Three-seed experiments on five scenes support the overall conclusion; training wall-clock time decreases by roughly 16%–21%, while rendering FPS slightly drops and rendering memory slightly increases.
Critical strengths
-
The matched-capacity experimental design is the strongest part of the paper. CAED and PWRS share identical Gaussian counts and PLY sizes in every scene (Table 1), cleanly isolating "early densification" as the single variable of interest and avoiding the apples-to-oranges capacity comparisons common in the compression literature. In Figure 1 the two methods share identical horizontal coordinates, making the fixed-capacity comparison explicit and faithful to the stated design intent.
-
The evaluation is honest and systematic. Paired t-tests are reported with 95% confidence intervals and p-values; three-seed experiments are included (Tables 6/10); ablations cover frontend components (Tables 4/5/14), schedule shape and budget amplitude (Tables 18/12), DoG interaction (Tables 15/16), and multiple capacity levels (Table 17). The appendix candidly discloses the three timing conventions (W/I/C), run provenance, and the fact that historical main runs did not preserve complete optimization arguments — a level of transparency uncommon in conference submissions.
Critical weaknesses
-
Limited novelty: the method is a combination of existing modules, and the headline claim is not supported by the ablations. All three frontend operators (edge score, long-axis split, square-root budget) come from ImprovedGS (acknowledged by the authors in the Figure 2 note), and the backend (RPS/SPS/DoG) comes entirely from PWRS. No new mechanism is proposed; the contribution is largely the empirical finding that "ImprovedGS-style densification in front of PWRS compression" helps. More importantly: (a) "compression-aware" in practice only means that the growth budget and the final target are configured separately — B(t) in Eq. (4) does not depend on N_tar at all, so the budget is not actually aware of the compression target; (b) the "Edge Densification" in the title is contradicted by Table 14: CAED vs. Gradient-LAS (which removes EAS) shows five-scene mean differences of ΔPSNR = −0.015 dB, ΔSSIM = +0.0001, ΔLPIPS = −0.0001 with a seed-mean std of 0.076 dB, and the paper itself concedes that EAS is "a sampling design choice rather than a consistently superior scoring signal." The component that names the paper contributes close to no measurable benefit, and this tension needs to be addressed head-on in the main text.
-
Attribution is not clean, effect sizes are small, and results are scene-dependent. The overall +0.2384 dB gain is driven by Mip-NeRF 360 and Tanks & Temples; on Deep Blending CAED is slightly worse, and in Table 17 playroom degrades by −0.2865 dB PSNR at the larger capacity. Table 4 shows that the budgeted frontend alone (B-G-I) already contributes most of the LPIPS improvement over PW ADC, while the EAS/LAS increments are small and inconsistent in direction. In addition, the abstract highlights quality gains and reduced training time but omits the rendering-side costs: FPS drops by 22.8 and peak memory rises by ~96 MiB — an important caveat for a "compact model" narrative.
-
Hyperparameter selection and statistical conventions raise soundness concerns. The appendix ("Randomness and coverage") admits that the historical per-scene selection of B_max "does not establish a tuning protocol entirely independent of test information." Since B_max varies per scene between 1.0M and 2.5M (Table 7), if these budgets were chosen with test data in the loop, the 13-scene gains may be subject to selection bias. Separately, the main text claims PSNR/SSIM/LPIPS improve in 11/13, 11/13, and 12/13 scenes, respectively, but per the appendix Table 8 values, PSNR decreases in four scenes (counter, bonsai, train, playroom) and SSIM decreases in three (bonsai, train, playroom), which does not reconcile with the stated counts — the authors should verify.
Questions for authors
-
Please reconcile Table 8 with the claimed "11/13, 11/13, and 12/13 scenes improve": per the appendix values, PSNR decreases on counter, bonsai, train, and playroom, and SSIM decreases on bonsai, train, and playroom. How do the counts align?
-
Table 14 shows that EAS yields essentially zero average gain over gradient-based sampling (with large variance): how much does EAS contribute to the 13-scene main results? Can you provide the 13-scene result for "budget + LAS without EAS" to support the "Edge" claim in the title?
-
B_max is selected per scene with no evidence that test information was excluded. Could you re-run the main experiments with B_max set as a function of N_tar (e.g., α·N_tar) and report sensitivity? If the growth budget is fully decoupled from the compression target, is "compression-aware" too strong a term?
-
In the DoG interaction ablation (Table 16), the interaction term is +0.64 dB on playroom but −0.15 dB on bicycle. How do you explain this scene-level reversal, and how much do the main results depend on DoG?