acceptodds
Under review as a conference paper at ICLR 2027

MotionAwareGS-SLAM: Temporal-Consistency-Aware Observation Integration for Dynamic Gaussian SLAM

Abstract

Observations useful for camera tracking and those suitable for persistent Gaussian mapping operate at fundamentally different temporal scales: a transient observation may provide a reliable pose constraint for the current frame, yet become temporally inconsistent when integrated into a persistent Gaussian map. Existing dynamic Gaussian SLAM methods mainly determine observation usability from current-frame reliability, implicitly assuming that observations beneficial for tracking are also suitable for persistent map integration. This mismatch can introduce temporally unstable information into long-term Gaussian representations even when camera tracking remains accurate. We propose MotionAwareGS-SLAM, a temporal-consistency-aware observation integration framework that incorporates temporal consistency evidence into dynamic Gaussian mapping. Specifically, we extract multi-frame temporal consistency evidence, regulate observation participation in persistent Gaussian updates, and perform asymmetric observation routing between camera tracking and persistent mapping. This design allows transient observations to contribute to robust tracking while preventing temporally unstable observations from degrading persistent Gaussian representations. Experiments on four TUM RGB-D dynamic sequences demonstrate the tracking-mapping reliability mismatch through independent future-only RGB-D and official-pose analysis, and validate the proposed observation routing through instrumented tracking and mapping pipelines. The results show that MotionAwareGS-SLAM improves persistent scene consistency while maintaining accurate camera trajectory estimation in dynamic environments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D Gaussian Splatting, dynamic SLAM, temporal consistency, observation integration, persistent Gaussian mapping

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZHMVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MotionAwareGS-SLAM, a temporal-consistency-aware observation integration framework for dynamic Gaussian SLAM. The method distinguishes observations useful for immediate camera tracking from those suitable for persistent Gaussian mapping. It employs frozen DINOv2 features, temporal consistency estimation, and asymmetric observation routing to regulate Gaussian creation, updates, and densification.

The evaluation includes four TUM RGB-D dynamic sequences, future-only geometric consistency analysis, camera trajectory measurements, and reconstruction quality comparisons. Although the authors report improvements in revealed-background reconstruction within a limited evaluation window, the overall experiments show mixed tracking performance and generally degraded full-frame reconstruction quality.

Critical strengths
  1. Relevant problem formulation. The distinction between short-term observation reliability for camera tracking and long-term compatibility with persistent map integration is conceptually meaningful. Treating these two requirements separately is a reasonable direction for dynamic Gaussian SLAM.

  2. Attempts to validate the integration mechanism. The paper introduces an independent future-consistency diagnostic and instruments Gaussian mapping operations to verify observation routing. The authors also report several negative experimental outcomes rather than presenting only favorable results.

Critical weaknesses
  1. Insufficient reconstruction quality and lack of convincing empirical improvements.

The most significant weakness is that the proposed method fails to demonstrate consistent improvements in reconstruction quality over its own baseline.

According to Table 3, Hard Commitment reduces SSIM on all four evaluated sequences and decreases PSNR on three of them. The Consensus variant also produces lower SSIM on every sequence and lower PSNR on three sequences.

Averaged over the four local evaluation sequences, the baseline achieves 16.634 dB PSNR and 0.6403 SSIM, whereas Hard Commitment achieves only 16.353 dB and 0.6232 SSIM, and Consensus achieves 16.376 dB and 0.6222 SSIM.

The tracking results provide little compensation for these reconstruction losses. Table 1 shows that Hard Commitment increases ATE RMSE on three of the four sequences, while Consensus also fails to achieve consistent improvements.

The authors emphasize a 7.977 dB improvement in revealed-background PSNR, but this result is obtained from a restricted evaluation window on a single sequence and does not demonstrate improved overall reconstruction quality. Furthermore, the qualitative results in Appendix D remain visually uneven, and the cross-method comparisons do not follow a fully matched evaluation protocol.

For a paper proposing an improved Gaussian SLAM framework, these results do not convincingly demonstrate a meaningful advance in reconstruction fidelity, tracking accuracy, or overall system performance.

  1. Limited methodological novelty, insufficient technical specification, and inadequate presentation of the architecture.

The proposed approach primarily combines existing components, including pretrained DINOv2 features, temporal feature comparison, heuristic consistency scoring, and observation filtering. The claimed methodological contribution is not sufficiently differentiated from existing dynamic SLAM strategies that already separate tracking support from mapping updates.

Moreover, the proposed tracking–mapping reliability mismatch is not directly established in terms of actual pose-estimation usefulness. As acknowledged in Section 3.2, current observation acceptance is only a proxy for tracking utility. Consequently, the diagnostic does not establish that an observation is beneficial for camera tracking while simultaneously harmful to mapping.

The routing mechanism also lacks sufficient operational specification. Although Eqs. (2)–(5) define temporal consistency scores and routing categories, the exact decision thresholds and the quantitative mapping from those scores to the three routing outcomes are not adequately explained.

The architecture and conceptual figures rely heavily on illustrative visual elements. The manuscript acknowledges AI assistance in producing illustrative figures. AI-assisted illustration is not itself a technical deficiency, but the resulting figures should provide precise algorithmic information rather than primarily serving as conceptual visualizations. In particular, the figures do not compensate for the missing operational details of the routing mechanism.

These issues substantially weaken the technical clarity, reproducibility, and overall scientific contribution.

Questions for authors
  1. Reconstruction quality: Given that both Hard Commitment and Consensus underperform the unweighted baseline in average PSNR and SSIM, what evidence demonstrates that the proposed method improves persistent reconstruction quality overall? Can the authors provide protocol-matched quantitative comparisons against competitive dynamic Gaussian SLAM methods across all evaluated sequences?

  2. Methodological specification: What are the exact thresholds and decision rules mapping the temporal consistency scores in Eqs. (2)–(3) to the three observation routes in Eq. (4)? How were the fixed coefficients selected, and how sensitive are the reconstruction results to these choices?

  3. Experimental validity: The all-map control in Table 4 has an ATE RMSE of 38.662 cm, substantially worse than the unweighted baseline reported for walking xyz in Table 1. What experimental differences explain this discrepancy? Can the claimed 7.977 dB revealed-background PSNR improvement be reproduced across multiple sequences using stronger, protocol-matched baselines?

Reviewer_12koOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper argues that "observations" in dynamic environments can be beneficial for tracking yet be unstable for persistent map integration. They propose a method in which temporal stability is used to allow transient (i.e dynamic) observations to be used in tracking, yet prevent them from degrading the map.

Critical strengths

The paper's motivation is interesting.

Critical weaknesses

The authors provide no quantitative comparison at all to any baseline or related work, including the two baselines shown in Fig. 1, besides the "external tracking context" in the appendix, which does not list their own results either. Since there is no comparison, this method cannot be evaluated. The authors provide no qualitative rendering results in the main text despite being a 3DGS SLAM method, the only one in the appendix is "Additional Mapping Visualization" showing their method vs only SplaTAM and Photo-SLAM w/ and w/o mask, which are non-dynamic SLAM methods, making the comparison unfair. The authors comparison Table I is between their own method and an "unweighted baseline", which was never specified anywhere in the text. Furthermore WildGS (referenced) has better ATE than those in Table I. e.g. fr3_walking/xyz 1.3 vs 1.538, fr3_walking/hs 1.6 vs 2.630. Thus the results in Table I undermine the claims. The authors only evaluate on 4 tum dynamic sequences, and the evaluation scope isn't convincing. Just as an example, no separation was made between high and low dynamic scenes (e.g. tum fr3 walking and static), Bonn dynamic sequences were not used. The paper reports "Hard" and "Consensus" results. Despite a careful reading, I could not tell which the author's method is. The paper is hard to read. The text is highly verbose but unclear, often with excess and out of context numerical results.

Questions for authors

I suggest the authors revise the manuscript to include both qualitative and quantitative comparisons to the relevant baselines, such as WildGS and DyPho-SLAM. For such papers, it is standard to evaluate on TUM and Bonn dynamic scenes, to test the method against a variety of dynamic objects (e.g. people, boxes, balloons), with different velocities and camera motions, while reporting ATE for pose estimation and PSNR/SSIM/LPIPS for mapping. It is also fairly standard to ablate the components used. Please follow these conventions and report results accordingly. The authors should address why their method performs worse on avg than their "unweighted baseline", which itself should be named and described in the text. Please provide fair comparisons to support the claims in this paper.

Reviewer_bpSwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MotionAwareGS-SLAM, which separates observation usage for camera tracking from admission into a persistent Gaussian map. Frozen visual features and temporal history provide consistency evidence, and observations are routed to tracking, mapping, or rejection. The motivation is reasonable, but the methodological description, experimental validation, and comparison with prior work are insufficiently developed to establish the contribution.

Critical strengths

Separating immediate tracking utility from suitability for persistent mapping is practically relevant. The authors also acknowledge performance trade-offs.

Critical weaknesses
  1. Insufficient methodological detail and validation. Temporal attention, cross-frame correspondence, and routing rules are underdefined. The main background-recovery experiment also involves substantially different trajectory errors, leaving the contribution of selective integration insufficiently isolated.
  2. Figure 1 does not clearly establish the methodological distinction. It does not explain how the proposed method and WildGS-SLAM would treat the same observation differently, leaving the novelty unclear. In addition, the illustrated input, pre-processing and rendered result are suspected generated by AI without verification, which are unacceptable. The static map still contains dynamic objects, i.e., human.
  3. Limited experimental scope. Evaluation covers only 4 sequences, with the strongest mapping result restricted to one short window. This provides insufficient scene diversity and evidence of generalization.
  4. Figure 3 is insufficiently informative. The trajectories largely overlap at the displayed scale. Enlarged regions with meaningful deviations or error-over-time plots would better reveal the differences.
  5. Insufficient related-work discussion. Coverage and technical comparison are limited.
Questions for authors

Please refer to above weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.