- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
PortGS addresses a narrow but real gap in transient-aware 3D Gaussian Splatting. The paper frames this as a representation-level coordination problem and proposes capacity–participation decoupling: each base Gaussian is frozen and paired with a trainable plastic counterpart; a learnable allocation fraction partitions the fixed base opacity between the pair, so plastic participation is a reallocated share of the retained reference rather than an independent opacity. Base and plastic are jointly rasterized in one standard 3DGS pass under the inherited transient-aware supervision. Validation covers NeRF On-the-go and RobustNeRF. Ablations isolate frozen vs. trainable base, independent vs. additive vs. paired participation, intervention timing, refinement duration, plastic scale, lower-tail view quality, and a pair-aware splitting variant for topology evolution.
Critical strengths
The key observation is well-motivated and cleanly instantiated. The observation that naively adding Gaussians to a converged transient-free representation grants the new capacity an independent opacity path is meaningful. Applying PortGS to three different backbones and reporting per-scene results across occlusion levels, plus a fixed-topology cross-backbone protocol to isolate the representation from backbone-specific densification, shows awareness that the claim is about a mechanism rather than a pipeline.
Critical weaknesses
The method is only evaluated as a late-stage post-hoc refinement, and the paper itself shows it is harmful if applied before ~20k steps, so the contribution is inherently a marginal add-on whose benefit depends on the backbone having converged to a sufficiently good anchor, a condition the paper does not characterize (e.g., how does anchor quality, or the fraction of occluded static evidence, predict the gain?). The absolute improvements are small: on RobustSplat, PortGS moves Corner from 26.50 to 26.79 PSNR and Patio from 21.70 to 21.90, and the "best overall" result is PortGS-on-DualSplat, i.e., the strongest backbone plus the add-on, rather than a demonstration that PortGS rescues a weak backbone. Meanwhile the topology-evolution results show the pair-aware splitting variant is roughly a wash (e.g., DS-Patio 22.076 → 22.074).
Questions for authors
(1) Can you provide a capacity-matched control: plastic primitives with independent opacity, same count, same κ, same fresh Adam optimizer, same fixed-topology protocol, compared against paired reallocation? If the gap shrinks substantially, how would you restate the contribution?
(2) How does the gain depend on anchor quality? Specifically, is there a measurable relationship between the PortGS improvement and (a) the anchor's PSNR, (b) the estimated fraction of occluded/incomplete static evidence, or (c) the backbone's residual artifact level? A characterization would convert "late-stage only" from a limitation into a usable deployment rule.
(3) For topology evolution, the pair-aware splitting variant is essentially neutral on DS-Patio. Do you have results under a native densification/pruning schedule (not the controlled variant)? If not, please state explicitly that compatibility with real schedules is untested, and temper the corresponding claim.