acceptodds
Withdrawn from ICLR 2027

PortGS: Paired Opacity Reallocation for Transient-Free 3D Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recovering static scenes from casually captured multi-view images with transient distractors remains challenging for 3D Gaussian Splatting (3DGS), as transient objects introduce cross-view inconsistencies while occluding static content, leaving the available static supervision incomplete. Existing transient-free 3DGS methods mainly suppress unreliable observations, disentangle scene components, or stabilize reconstruction. However, even after transient-aware training has established a reasonably reliable static representation, occlusion-induced missing evidence can leave it under-constrained, making further refinement from reliable but incomplete observations a distinct representation-level challenge. To address this problem, we propose PortGS, which coordinates representation retention and refinement through capacity-participation decoupling, expanding learnable capacity without automatically granting it independent rendering participation. PortGS organizes the scene into persistent base-plastic Gaussian pairs, with base primitives retaining established scene parameters and plastic primitives providing additional geometric and appearance capacity. We further introduce base-conditioned paired opacity reallocation, in which a trainable allocation variable partitions the base opacity reference between the paired primitives. This makes plastic rendering participation relative to the retained base rather than independently parameterized. Base and plastic are then jointly rasterized as a unified static representation under the inherited transient-aware supervision.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Novel View Synthesis, Transient-Free Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4xUkOpenReview ↗
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.

Reviewer_cdf7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PortGS, a method that provides additional refinement capacity to Gaussian representation by freezing base Gaussians and introducing trainable plastic Gaussians, each paired with its frozen base counterpart. The method aims to address under-constrained reconstruction due to occlusion-induced missing evidence. It reallocates opacity between base and plastic Gaussians to control their rendering contributions. The paper shows performance gains when its method is integrated into several existing distractor-free 3DGS methods as a plug-in.

Critical strengths
  • The proposed method shows significant performance gains when combined with DualSplat, and shows moderate gains with DeSplat and RobustSplat.
  • The proposed method is relatively simple and effective, and the paper is generally easy to follow.
Critical weaknesses
  • The connection between the motivation and the proposed method is not sufficiently convincing. While the paper highlights underconstrained reconstruction due to occlusion-induced missing evidence, the method instead increases refinement capacity of representation by introducing additional plastic Gaussians while relying on the same incomplete observations. It is unclear how this specifically addresses the stated problem.
  • No analysis or discussion is given on why the proposed method shows substantially larger performance gains on DualSplat than on DeSplat or RobustSplat.

for minor concerns:

  • In some scenes, the performance improvements over in-place continuation appear marginal, despite doubling the number of Gaussians through base/plastic representation.
  • Fig. 1 'Fixed pairwise opacity total' part is potentially misleading as it depicts changes in Gaussian size rather than opacity.
Questions for authors
  • Could you provide ablation results on additional scenes or report the average performance across all scenes?
  • Could you provide a comparison of training time, memory, and model storage size for the baseline methods with and without PortGS?
Reviewer_MHABOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PortGS introduces capacity-participation decoupling for transient-free 3D Gaussian Splatting through persistent base-plastic Gaussian pairs and base-conditioned paired opacity reallocation. Base primitives retain established scene parameters, while plastic primitives provide trainable geometry and appearance, with a learnable allocation fraction partitioning the base opacity reference between each pair. Experiments on NeRF On-the-go and RobustNeRF demonstrate PSNR/SSIM improvements when integrated into DeSplat, RobustSplat, and DualSplat.

Critical strengths

PortGS introduces a lightweight conceptual modification of paired opacity reallocation, without requiring an additional transient-prediction module or a custom renderer. It consistently improves PSNR/SSIM across two benchmarks on three underlying methods, with particularly substantial gains on DualSplat, demonstrating its potential as a practical refinement strategy for existing transient-aware 3DGS pipelines.

Critical weaknesses
  • Insufficient evaluation for ablation studies: Table 4 only covers Corner and Patio, the two scenes with the smallest gains for DualSplat. Ablations on high-gain scenes from both datasets are needed to determine whether the larger improvements can be achieved through in-place continuation alone or require the proposed paired opacity reallocation. Moreover, it is unclear whether “Independent opacity” optimizes both base and plastic opacities. This control is important for isolating the benefit of the fixed pairwise opacity budget.

  • Limited technical novelty: The technical contribution appears relatively incremental, consisting of a frozen base representation, trainable Gaussian copies, and a shared scalar that partitions each pair’s opacity. The coupled refinement signal follows directly from this parameterization through standard differentiable rasterization. The paper should more clearly establish its conceptual advance over LapisGS [1] and Augmented Radiance Field [2], both of which preserve existing Gaussian geometry and appearance while optimizing additional Gaussians and adjusting the original opacities.

[1] Shi et. al., 3DV 2025, LapisGS: Layered Progressive 3D Gaussian Splatting for Adaptive Streaming. [2] Yang et. al. ICLR 2026, Augmented Radiance Field: A General Framework for Enhanced Gaussian Splatting.

Questions for authors
  • Does the “Independent opacity” ablation optimize both base and plastic opacities while freezing the base geometry and appearance? If not, could the authors provide this control to isolate the benefit of the fixed pairwise opacity budget?

  • Could the authors report the key ablations across all scenes in Tables 1 and 2 to determine whether paired opacity reallocation also shows larger advantages over the ablated variants on high-gain scenes (e.g., Spot, Statue, Yoda)?

  • Beyond persistent base-plastic Gaussian pairs and base-conditioned paired opacity reallocation, what distinguishes PortGS from LapisGS and Augmented Radiance Field, which already preserve existing Gaussian geometry and appearance while optimizing additional Gaussians and adjusting the original opacities?

  • Why does PortGS yield substantially larger gains on DualSplat than on DeSplat and RobustSplat, and what aspects of the underlying methods account for this difference?

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

Related papers

Discussion (0)

Sign in to comment.