acceptodds
Under review as a conference paper at ICLR 2027

SAGE-GS: Spatially Weighted Event Supervision and Sparse Temporal Gating for Dynamic Gaussian Splatting

Abstract

Event observations provide inter-frame brightness-change constraints for dynamic scene reconstruction, but their spatial distribution does not necessarily reflect where supervision is most useful. In addition to geometric deformation, dynamic Gaussians need to modulate their rendering contributions over time to accommodate both persistent and temporally localized contributions. We propose SAGE-GS, which improves event-assisted dynamic Gaussian reconstruction through event supervision allocation and temporal representation and optimization. First, we use CTA to fuse current RGB and event features, incorporating dynamic-region priors, soft event-density weights, and amplitude-based confidence to predict spatial weights for event residuals. Second, we equip each dynamic Gaussian with a learnable temporal center, a fixed-scale temporal kernel, and an individual gating strength. Zero initialization, interval projection, and sparsity regularization enable selective temporal suppression while allowing Gaussians that do not require suppression to retain their original opacity. Both branches are jointly trained using RGB and weighted event reconstruction objectives, while the deformation network models geometric motion. Experiments on eight dynamic scenes show that SAGE-GS improves average PSNR by approximately 0.44 dB over E-D3DGS.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic scene reconstruction, Gaussian splatting, RGB–event fusion, Adaptive supervision weighting, Sparse temporal gating

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_r8smOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAGE-GS, an event-guided dynamic Gaussian reconstruction framework based on E-D3DGS. The method introduces two modules: spatial event-supervised weighting, which adaptively adjusts the weights of event reconstruction errors by fusing RGB-Event features and spatial priors; and sparse temporal opacity gating, which achieves time-dependent opacity adjustment by modulating the rendering contribution of the dynamic Gaussian at different time steps. Experiments on 8 synthetic scenes and 4 real-world scenes show that, compared to E-D3DGS, this method achieves average PSNR improvements of 0.52 dB and 0.24 dB, respectively.

Critical strengths

This paper addresses a problem of practical research significance: event signals do not always reliably reflect the motion of objects in the real world. By fusing features from RGB and event data and combining them with spatial prior information to perform adaptive weight adjustment for event-based supervision, this paper proposes a reasonable and well-motivated technical solution.

This paper conducts experimental evaluations on both synthetic and real-world scenes and provides ablation experiments for each module. The reported experimental results show that, compared to the baseline method E-D3DGS, this approach achieves performance improvements across multiple test scenes, preliminarily validating the effectiveness of the proposed modules.

Critical weaknesses

This method primarily combines event loss weighting with temporal opacity gating. The RGB-Event feature fusion draws on CompEvent, while a similar temporal gating mechanism has been studied in previous dynamic Gaussian methods. The paper does not fully explain the technical innovation or necessity of this combination.

Compared to using supervised weighting alone, temporal gating yields only an additional average PSNR improvement of approximately 0.08 dB on synthetic scenes, and performance actually declines in some scenes. Detailed weight ablation experiments were conducted only on the “fan” scene, and the study lacks experiments with multiple random seeds as well as robustness tests under event noise and lighting variations. Furthermore, the visual evidence provided in Figures 3 and 4 is limited and does not sufficiently support the reported performance improvements.

Questions for authors

Q1. In Eqs. (3)–(5), the weight predictor uses reconstruction residuals as input. How do the authors rule out the possibility that it simply learns to down-weight difficult pixels? Please compare against a predictor without residual input and analyze the relationship between predicted weights and reconstruction errors.

Q2. Adding temporal gating improves average synthetic PSNR by only about 0.08 dB over supervision weighting alone, with regressions on some scenes. Could the authors report mean and standard deviation over multiple random seeds on representative scenes, together with results at held-out timestamps? A comparison with fixed temporal gating would also help establish the value of the learned gates.

Q3. Why is the detailed weighting ablation conducted only on fan? Please extend it to at least one additional synthetic scene and one real scene. A limited robustness test under event noise or illumination changes, together with enlarged rendering crops and error maps, would better support the main claims.

Q4. In Table 1, the T-only SSIM on celestial_globe is 0.997, while the S-only and T-only results on Eagle are identical. Could the authors verify these entries and clarify how the best-PSNR checkpoints were selected? Please also report the main hyperparameters and the computational overhead of the added modules.

Reviewer_j3YMOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SAGE-GS, an extension of E-D3DGS for event-guided dynamic scene reconstruction. It introduces two components: spatial weighting of event supervision and sparse temporal opacity gating. The spatial branch predicts weights from fused RGB–event features and reconstruction residuals, combining them with event-response weights, a dynamic-region prior, and amplitude reweighting. The temporal branch learns a temporal center and suppression strength for each dynamic Gaussian, with sparsity regularization encouraging selective opacity modulation. The method is evaluated on eight synthetic and four real-world scenes, with baseline comparisons, component ablations, and gating analyses. The authors report average PSNR improvements of approximately 0.52 dB and 0.24 dB on synthetic and real-world scenes, respectively, over the E-D3DGS baselines used in the paper.

Critical strengths

This paper studies a meaningful problem in event-based dynamic scene reconstruction. The authors combine RGB and event information to adjust supervision weights and introduce temporal opacity gating. Experiments on synthetic and real scenes, together with ablation studies, show some benefits of the proposed method.

Critical weaknesses

Some key implementation details are not provided, including the encoder architectures, feature dimensions, GRU settings, prediction head, and the calculation of the event-supervision target. Many hyperparameter values and training settings are also missing. It is difficult to reproduce the method from the current description.

The reasons for choosing these weighting functions and multiplying them together are not sufficiently explained. The detailed weighting ablations are only conducted on the fan scene, and the contribution of the amplitude-reweighting factor is not tested separately. In addition, the reproduced E-D3DGS baseline on real scenes is about 2.11 dB lower than the original results, while the reported improvement is only 0.24 dB. This large baseline difference needs a clear explanation before the performance claims can be properly assessed.

Questions for authors
  1. Could the authors provide the detailed network structures, feature dimensions, GRU settings, the calculation of E_{k,e}(u), and the complete hyperparameter values and training settings?

  2. Why are these particular weighting functions used, and why are they multiplied together? What experiments support the combination of w_density and w_conf, which have opposite effects on event magnitude? Do the conclusions also hold on other scenes?

  3. Why is the reproduced E-D3DGS baseline about 2.11 dB lower than the original results on real scenes?

Reviewer_DHF5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In my view, SAGE-GS builds upon Event-D3DGS, aiming to reconstruct dynamic scenes by leveraging sparse RGB frames and event observations. Its spatial branch incorporates learned RGB-event residual weights, factored by event responses, dynamic regions, and magnitudes, to redistribute event-based supervision signals. Its temporal branch adjusts the opacity of dynamic Gaussian primitives by utilizing the temporal center of each primitive alongside a suppression intensity constrained by sparsity regularization. When evaluated across multiple datasets, this method demonstrates performance improvements over Event-D3DGS.

Critical strengths
  1. This method proposes solutions for two key aspects of event-guided reconstruction—the allocation of spatial supervision and the temporal control of Gaussian contributions—with experimental results confirming their effectiveness.
  2. The paper transparently reports scene-dependent regressions and distinguishes temporal suppression sparsity from Gaussian-count reduction or rendering acceleration.
  3. The paper is well-written, and its core ideas are easy to grasp.
Critical weaknesses
  1. The baseline is considerably outdated. Event-D3DGS was published at CVPR 2025, with work completed in the fall of 2024. More recent methods relevant to this task have since emerged; for instance, DEGS [1] addresses the closely related problem of RGB–event dynamic Gaussian reconstruction, while E2EGS [2] and eRetinexGS [3] (Different task, but relatively same idea) offer pertinent precedents for spatially weighted or adaptive event-based supervision.
  2. The temporal mixing function is algebraically equivalent to the persistence-gating form in FreeTimeGS++. The benefits of the proposed initialization, projection, and sparsity regularization are not isolated experimentally. ( FreeTimeGS++: \phi_i(t)=p_i+(1-p_i)G_i(t) v.s. SAGE-GS: h_i(t)=(1-\alpha_i)+\alpha_iG_i(t) ) Given that non-uniform spatial weighting of event-based supervision is already a well-established concept in tasks involving event cameras, this approach is better characterized as an incremental development rather than one featuring significant innovation.
  3. The spatial branch lacks comparisons against simpler weighting and fusion alternatives. Moreover, the small, scene-dependent gating gains are not supported by repeated-run or dedicated temporal-interpolation evaluations.

Reference: [1] DEGS: Deformable Event-based 3D Gaussian Splatting from RGB and Event Stream, TVCG 2025. [2] E2EGS: Event-to-Edge Gaussian Splatting for Pose-Free 3D Reconstruction, CVPR 2026. [3] eRetinexGS: Retinex Modeling for Low-Light Scene Enhancement via Event Streams and 3D Gaussian Splatting, CVPR 2026.

Questions for authors
  1. Could you compare the learned weighting against simple residual-based weighting and ordinary RGB–event feature concatenation on multiple scenes? How do the resulting weights behave under controlled event noise and illumination changes?
  2. What benefits specifically come from zero initialization, projected gate strengths, and sparsity regularization compared with an existing sigmoid-based persistence gate or a gate without sparsity regularization? Please report repeated-run results, particularly for scenes where gating hurts.
  3. Could you evaluate dynamic-region reconstruction and temporal consistency, including times between observed RGB frames? These evaluations would more directly test the motivation for event supervision and temporal opacity control.
  4. Please also specify the temporal width, center initialization, weighting hyperparameters, network architecture, and training/rendering costs for reproducibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.