acceptodds
Under review as a conference paper at ICLR 2027

T-FastGS: Transient-Robust and Efficient 3D Gaussian Splatting

Abstract

Transient objects break the multi-view consistency assumption of 3D Gaussian Splatting (3DGS), leading to degraded reconstruction quality. Existing methods struggle to yield robust reconstruction in the presence of transients while preserving high training efficiency. To tackle this challenge, we present T-FastGS, a transient-robust and efficient reconstruction framework. We find that transient content introduces ambiguity into error-aware Gaussian density control. To mitigate this problem, we propose Mask-Guided Error-Aware Density Control (MGDC), which constructs reliable cues for density adjustment by distinguishing static and transient regions. Moreover, we identify a timescale mismatch between transient mask learning and Gaussian optimization. Motivated by this finding, we introduce an Asynchronous Alternating Optimization (AAO) scheme that cuts redundant updates to the mask predictor and retains training efficiency. Experiments show that T-FastGS achieves competitive rendering quality with state-of-the-art transient-free 3DGS methods on two benchmark datasets, while delivering around 3× or higher training speedup over these methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Transient Object Removal, Robust Reconstruction, Efficient Training

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_NyM5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary
  1. T-FastGS extends FastGS to reconstruct clean static Gaussian scenes from multi-view images containing transient objects, targeting both reconstruction robustness and training efficiency.
  2. Its Mask-Guided Error-Aware Density Control (MGDC) separates reliable static errors from transient interference, preserving densification opportunities in transient regions while excluding them from pruning-score computation.
  3. Asynchronous Alternating Optimization (AAO) updates the mask predictor less frequently and reuses cached masks, reducing overhead while Gaussian optimization continues.
  4. Experiments on NeRF On-the-go and RobustNeRF report rendering quality comparable to existing transient-robust methods, with substantially fewer Gaussians and approximately 3–3.5× faster training than RobustSplat. Ablations evaluate the contributions of MGDC and AAO.
Critical strengths
  1. Mask-based training suppresses transient interference, while AAO reduces the cost of maintaining these masks by avoiding redundant updates. Its ablation shows a 38.9% reduction in T-FastGS training time with negligible changes in rendering quality (Table 4).
  2. MGDC thoughtfully integrates masks into FastGS's density-control mechanism. Its asymmetric design preserves densification opportunities for occluded static structures while excluding transient errors from pruning scores, addressing a limitation of simply masking the reconstruction loss (Sec. 3.3 and Table 3).
  3. The method achieves a strong quality–efficiency trade-off across two benchmarks. On NeRF On-the-go, it delivers rendering quality comparable to RobustSplat with approximately 3.5× faster training and 0.20M rather than 0.98M Gaussians (Appendix Table 9).
Critical weaknesses
  1. The MGDC ablations do not isolate the contributions of its densification and pruning strategies or establish the benefit of their asymmetric combination. Separate component ablations, as provided in FastGS, are missing (Tables 3 and 5).
  2. AAO lacks an ablation of mask-update frequency. The nearly unchanged rendering quality with and without AAO in Table 4 leaves the choice of update interval and the quality–efficiency trade-off across different frequencies insufficiently justified.
  3. The reported training time may exclude DINOv2 feature precomputation for the input images, which occurs before timing begins in the supplied code. This leaves the end-to-end runtime and fairness of the reported speedup unclear.
Questions for authors
  1. Could you ablate different mask-update intervals in AAO, reporting rendering quality and training time to justify the chosen frequency and demonstrate the effectiveness of the schedule?
  2. Could you separately ablate MGDC's densification and pruning strategies, including their combination, to quantify each component's contribution to rendering quality, Gaussian count, and training efficiency?
  3. Could you report the DINOv2 feature-precomputation time and total runtime including preprocessing, using a consistent timing protocol for all methods?
  4. I would be willing to raise my score if the authors satisfactorily address these concerns.
Reviewer_R5ckOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a new method named T-FastGS, which helps reconstruct static scenes from multi-view images containing transient distractors while preserving the rendering quality and training efficiency. Experiments on NeRF On-the-go and RobustNeRF show competitive reconstruction quality with substantially faster training than existing transient-robust 3DGS methods.

Critical strengths
  1. Practical efficiency. The framework achieves a useful quality-efficiency trade-off, substantially reducing training time and Gaussian count while retaining competitive rendering quality.

  2. Relevant ablations. The paper compares multiple density-control variants, tests AAO on both T-FastGS and RobustSplat, and reports repeated-run statistics for these ablations.

Critical weaknesses
  1. Limited novelty. Feature-based and residual-based transient masking is already established in SpotLessSplats (Sabour et al., 2025) and RobustSplat (Fu et al., 2025), while the efficient backbone is inherited from FastGS. The main additions are an asymmetric indicator-map heuristic and a periodic update schedule, making the contribution primarily an incremental integration of existing techniques.

  2. Modest peformance MGDC gains. MGDC improves PSNR by only 0.02-0.19dB across Tables 3 and 5 while consistently using more Gaussians. A matched-capacity comparison is needed to distinguish better allocation from additional representation capacity and establish the incremental value beyond FastGS with transient masking.

  3. Limited visualization of MGDC masks. Beyond the examples in Fig. 2, the paper lacks broader visualizations of MGDC’s indicator masks across scenes and training stages, making it difficult to assess their reliability.

  4. Overly fragmented method writing. The method section contains too many equations for simple operations, which should be consolidated or omitted to improve readability so as to emphasize the core contributions.

Questions for authors
  1. Does MGDC still outperform the Global Error baseline when both methods use the same number of Gaussians?

  2. How does the method recover static regions mistakenly classified as transient during early optimization, given that masking suppresses the reconstruction supervision needed to correct those regions?

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

This paper proposes T-FastGS, a transient-robust and efficient 3D Gaussian Splatting framework built upon FastGS. It introduces Mask-Guided Error-Aware Density Control (MGDC) to improve Gaussian densification and pruning in the presence of transient objects, and Asynchronous Alternating Optimization (AAO) to reduce the computational overhead of mask prediction. Experiments demonstrate competitive reconstruction quality with improved training efficiency. However, the contributions are primarily engineering-oriented, with limited methodological novelty and insufficient new insights.

Critical strengths
  1. Well-motivated problem formulation. The paper addresses a practical challenge in 3D Gaussian Splatting: achieving both transient robustness and efficient scene reconstruction. The authors identify that transient objects introduce ambiguity into error-aware Gaussian density control, potentially leading to inappropriate densification and pruning. This observation provides a reasonable motivation for the proposed method.

  2. Reasonable experimental evaluation. The authors evaluate their method on two transient-robust reconstruction benchmarks, NeRF On-the-go and RobustNeRF, with comparisons against several existing methods. The experiments include quantitative evaluations, qualitative visualizations, runtime measurements, and ablation studies of the proposed components.

Critical weaknesses
  1. Limited improvements and insufficiently controlled MGDC ablations. The quantitative improvements brought by MGDC are relatively modest. For example, Table 3 reports PSNR gains of only 0.19 dB on Mountain and 0.13 dB on Spot compared with the Global Error baseline. Meanwhile, MGDC consistently produces more Gaussian primitives than the compared allocation strategies. Therefore, it remains unclear whether the improvements originate from more effective density control or simply from increased model capacity. The authors should provide comparisons under matched Gaussian counts to better isolate the contribution of the proposed MGDC mechanism.

  2. Limited novelty of Asynchronous Alternating Optimization (AAO). The proposed AAO mechanism primarily reduces computational overhead by periodically updating the transient mask predictor while continuously optimizing Gaussian parameters. Although this strategy is effective in reducing training time, periodic parameter updates and alternating optimization are well-established techniques. The proposed method does not appear to introduce a fundamentally new optimization formulation or adaptive scheduling strategy. Furthermore, the choice of the update interval KK is not sufficiently justified. Additional comparisons with alternative fixed-frequency and adaptive update schedules would help establish the necessity and effectiveness of the proposed design.

  3. Insufficient attribution of the reported training speedup. The paper emphasizes approximately 3× training acceleration over existing transient-robust 3DGS methods. However, the proposed framework is built upon FastGS, which already provides substantial training acceleration through efficient density control and rasterization. According to Table 7, T-FastGS actually requires more training time than vanilla FastGS due to the additional transient mask prediction modules. Therefore, much of the reported speedup over previous transient-robust methods may originate from the inherited FastGS framework rather than the proposed components. The authors should more clearly distinguish the efficiency gains inherited from FastGS from those introduced by MGDC and AAO.

  4. Lack of targeted evaluation of transient removal and occluded static reconstruction. The proposed method aims to improve reconstruction robustness in the presence of transient objects, particularly by preserving densification opportunities for occluded static structures. However, the evaluation primarily relies on global image reconstruction metrics, which do not directly measure the effectiveness of transient removal or background recovery. The authors should consider region-specific evaluations, such as reconstruction quality in previously occluded static regions, or experiments with controlled transient occlusion ratios. Such analyses would provide stronger evidence that MGDC addresses the intended problem.

Questions for authors

please refer to weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.