acceptodds
Under review as a conference paper at ICLR 2027

TAG4D-SLAM: Track-Anchored 4D Gaussian SLAM with Feed-Forward Priors

Abstract

Dynamic scenes remain a fundamental challenge for visual SLAM: moving objects violate the static-world assumption, while simply filtering them out discards valuable scene content. To address this, we present TAG4D-SLAM, a monocular dynamic Gaussian SLAM framework that jointly performs accurate camera tracking and high-fidelity 4D mapping with priors from the feed-forward reconstruction model. Specifically, we augment the feed-forward model with dynamic prediction, providing pose, depth, confidence, Gaussians, and dynamic probability for SLAM. Going beyond its explicit predictions, we further exploit the latent cross-view geometry encoded in internal attention to perform correspondence-aware patch sampling (CAPS) for camera tracking, turning implicit geometric knowledge into reliable constraints for bundle adjustment. For static mapping, prior Gaussians are compacted through voxelization and incrementally consolidated into a persistent map according to scene coverage. For dynamic content, rather than rejecting moving observations or relying on frame-local optical flow propagation and learned deformation fields, we introduce Track-Anchored Gaussians (TAGs) that anchor canonical dynamic Gaussians to long-lived 3D motion tracks. This decouples persistent scene content from temporal motion and establishes a persistent motion scaffold across frames, enabling more stable temporal association under large motion and occlusion. The resulting system yields accurate dynamic perception, robust camera tracking, and high-quality 4D reconstruction from monocular RGB input. Experiments on challenging dynamic datasets demonstrate our method's state-of-the-art performance in both camera tracking and 4D reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic SLAM, 3D Gaussian Splatting, Feed-forward Priors

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8VaLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper takes monocular RGB video and uses a pretrained WorldMirror model to provide geometry and Gaussian priors. Its camera tracker favors static patches selected using the model’s cross-frame attention. For mapping, it combines static Gaussians across frames and moves dynamic Gaussians with 3D trajectories obtained from tracked image points. The experiments cover motion segmentation, camera trajectories, rendering, and runtime. I find the rendering results on Bonn more persuasive than the tracking results: the method produces cleaner examples of moving objects, while its average trajectory errors are only slightly below the strongest monocular baselines.

Critical strengths
  1. Maintaining moving objects in a monocular SLAM map is useful, and the Bonn rendering examples show fewer visible artifacts around people and moving objects. Table 5 reports 27.28 dB average PSNR against 26.11 dB for Flow4DGS-SLAM. This is encouraging, particularly given the RGB input, though I would describe the full-image gain as moderate.

  2. The CAPS ablation gives a reasonably direct check on the tracking component: replacing its sampling strategy raises Bonn ATE from 2.22 to 2.59 cm. The result makes the attention-based patch choice more credible, even though the final improvement over other complete trackers is small.

Critical weaknesses
  1. The tracking advantage is very small. Average Bonn ATE is 2.22 cm versus 2.30 cm for DROID-W; on Wild-SLAM it is 0.56 cm versus 0.66 cm for the best listed monocular alternatives. Wild-SLAM novel-view PSNR improves from 20.31 to 20.56 dB over Flow4DGS-SLAM. Without repeated runs or an uncertainty estimate, it is difficult to judge how much weight to place on these differences.

  2. The evidence does not yet establish that anchoring Gaussians to long-lived tracks is responsible for the claimed robustness to motion and occlusion. The “without TAGs” ablation removes dynamic mapping altogether, rather than comparing against another way to move dynamic Gaussians within the same pipeline. Dynamic rendering is evaluated, but track survival, recovery after occlusion, and held-out dynamic views or times are not quantified. TrackerSplat also uses point tracks to guide Gaussian motion in a different setting; discussing that relationship would help clarify this paper’s contribution.

Questions for authors
  1. Could you compare TAGs with another motion-aware Gaussian update while keeping the same WorldMirror priors, point tracker, and optimization budget? How does the proposed anchoring differ from TrackerSplat’s use of point tracks?

  2. How were the dynamic-only images and masks in Table 10 generated for each method, and were those frames used during map optimization?

3 comments
AuthorsOpenReview ↗

Respomse to Reviewer 8VaL (1/2)

We sincerely thank the reviewer for the constructive feedback and for recognizing the benefits of our monocular dynamic reconstruction framework, particularly the improved rendering of moving objects and the effectiveness of CAPS. Our point-by-point responses are provided below.

W1: Magnitude and repeatability of the tracking and rendering gains

A1: Table R3-1 reports the mean and standard deviation of dataset-average ATE over five runs for DROID-W, π³mos-SLAM, and our method. On Bonn the three results are 2.30/0.00924, 2.37/0.186, and 2.22/0.01747 cm. On Wild-SLAM MoCap they are 0.96/0.00431, 0.66/0.1068, and 0.56/0.00266 cm. Our standard deviations remain 0.01747 cm and 0.00266 cm, while the standard deviation of π³mos-SLAM reaches 0.186 cm on Bonn and 0.1068 cm on Wild-SLAM MoCap, about 11× and 40× larger than ours. Across these repeated runs, our method is stably better than the other two baselines.

Table R3-1. Mean and standard deviation of dataset-average ATE over five runs(cm)

DROID-Wπ³mos-SLAMOurs
Bonn Avg./Std.2.30/0.009242.37/0.1862.22/0.01747
Wild-SLAM MoCap Avg./Std.0.96/0.004310.66/0.10680.56/0.00266

For novel-view synthesis, ours achieves better rendering quality using only monocular RGB, without ground-truth depth. As reported in Table 4, our method achieves higher PSNR on all five sequences and lower average LPIPS than Flow4DGS-SLAM, which uses RGB-D input. Table 6 further shows that our system runs at 0.94 FPS, compared with 0.36 FPS for Flow4DGS-SLAM, providing approximately 2.6× higher throughput. Our method thus combines improved rendering quality with faster processing and reduced input requirements. The standard deviation of our average PSNR across five runs is 0.005 dB, supporting the repeatability of these results. Table R3-2 provides the per-sequence values.

Table R3-2. Standard deviation of novel-view PSNR across five runs on Wild-SLAM MoCap, measured in decibels.

ScenePersonRacketTable1Table2UmbrellaAvg.
Std.0.0040.0030.0110.0170.0070.005
AuthorsOpenReview ↗

Respomse to Reviewer 8VaL (1/2)

W2 & Q1. Contribution of TAGs and Comparison with Alternative Dynamic Representations

A2: We acknowledge that TrackerSplat is an important related work deserving further discussion. Both methods leverage point tracking to improve dynamic Gaussian reconstruction, but they differ substantially in their problem settings and motion modeling strategies.

TrackerSplat primarily addresses dynamic Gaussian reconstruction from calibrated multi-view observations. It uses 2D pixel trajectories to estimate Gaussian motion in image space, then recovers their 3D centers and covariance matrices through multi-view geometric constraints, followed by Gaussian parameter optimization.

In contrast, TAG4D-SLAM targets online monocular dynamic SLAM with unknown camera poses and without ground-truth depth. We recover long-lived 3D trajectories from 2D point tracking, estimated depth, and optimized camera poses. Canonical Gaussians are explicitly associated with these trajectories and propagated directly according to their 3D displacements. These associations remain fixed throughout incremental map updates, while coverage-aware Gaussian insertion integrates newly observed dynamic content.

To further validate the contribution of our motion representation, we conduct additional controlled ablations by replacing TAGs with TrackerSplat's dynamic motion modeling mechanism and a control-point-based MLP deformation model adopted from 4DGS-SLAM, while keeping the remaining components unchanged. As shown in Table R3-3, TAGs achieve 27.28 dB PSNR, outperforming the TrackerSplat-based and control-point-based alternatives by 2.44 dB and 1.45 dB, respectively. TAGs also achieve higher SSIM and lower LPIPS, demonstrating the effectiveness of our track-anchored dynamic representation.

Table R3-3. Ablation study comparing different dynamic motion representations.

MethodPSNR ↑SSIM ↑LPIPS ↓
w/ TrackerSplat Motion Model24.840.8690.199
w/ Control Points + MLP25.830.8810.189
w/ TAGs (Ours)27.280.9020.156

Q2. Dynamic-Only Rendering Evaluation Protocol in Table 10

A3: For the ground-truth dynamic-only images, we apply SAM3 to segment dynamic objects in the original RGB frames and retain only the corresponding dynamic regions. For all compared methods, static and dynamic Gaussians are represented separately. We therefore render only the dynamic Gaussian components to obtain the dynamic-only reconstruction images, ensuring that the evaluation focuses specifically on dynamic scene content. Regarding the evaluation frames, all frames are available to the system during online SLAM and may be used for incremental map optimization. Therefore, Table 10 evaluates the dynamic reconstruction quality within the online mapping setting, rather than generalization to held-out frames or unseen timestamps.

Dear Reviewer 8VaL,

Thank you again for taking the time to review our paper and share your comments. We hope our earlier response has addressed your concerns. Since the ICLR discussion period ends soon, we would greatly appreciate your feedback on our response, and we are happy to answer any further questions.

Best regards,

The Authors

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

TAG4D-SLAM reconstructs static and dynamic scenes from monocular RGB video using feed-forward geometric priors. Its main designs include attention-guided patch sampling for camera tracking and Track-Anchored Gaussians, which separate persistent scene content from motion through long-lived 3D tracks. Evaluations on DAVIS, Bonn, and Wild-SLAM MoCap report improved segmentation, tracking, and rendering over evaluated baselines, with ablations supporting the main components.

Critical strengths
  1. Persistent dynamic reconstruction from monocular RGB. Track-Anchored Gaussians separate canonical scene content from motion through long-lived 3D tracks. The reported reconstruction gains over RGB-D baselines are substantial.

  2. Effective use of internal attention for camera tracking. CAPS converts cross-view attention selectivity into a patch-sampling prior for bundle adjustment. Removing CAPS increases Bonn ATE from 2.22 to 2.59 cm, while additional comparisons favor it over gradient- and SIFT-based sampling.

Critical weaknesses
  1. Anchor-based Gaussian representations already exist in works such as Anchored 4DGS, SC-GS, and 4DGS-SLAM. The later two use sparse motion controls to drive dynamic Gaussians. TAGs specifically use long-lived 3D tracks as anchors, but the ablation replaces TAGs with a static map rather than an alternative dynamic representation under matched inputs, leaving the advantage of this anchoring design unisolated.

  2. The claimed robustness to occlusion and long-term temporal association is not directly quantified. Evaluation mainly reports camera trajectory error and per-frame rendering metrics, which do not establish persistent dynamic correspondence or accurate recovery after occlusion.

Questions for authors
  1. Novelty and attribution. How does TAG anchoring differ from the sparse motion controls in SC-GS and 4DGS-SLAM beyond using externally estimated long-lived tracks?

  2. Track loss and reappearance. What precisely determines whether a track provides reliable support, and what happens to its bound Gaussians when support is lost? Upon reappearance, are the original track identity and Gaussian associations recovered, or are new tracks and Gaussians created? How does the system prevent duplicate content or incorrect associations?

  3. Propagation of tracking and depth errors. The 3D anchors depend on 2D tracks, predicted depth, and estimated camera poses, while tracks and Gaussian-to-track associations remain fixed during map refinement. Does the system subsequently correct erroneous anchor positions or bindings? What evidence indicates how these errors affect reconstruction under large motion or prolonged occlusion?

  4. Evidence for temporal persistence. Beyond per-frame rendering metrics and qualitative comparisons, do the existing evaluations measure dynamic trajectory accuracy, temporal consistency, or recovery after occlusion?

2 comments
AuthorsOpenReview ↗

Respomse to Reviewer PBRm (1/2)

We sincerely thank the reviewer for the constructive and insightful feedback, as well as for recognizing the effectiveness of our monocular dynamic reconstruction framework and correspondence-aware patch sampling (CAPS). Our point-by-point responses are provided below.

Q1&W1. Novelty of TAGs and comparison with alternative dynamic representations

A1: SC-GS employs learnable control points, whose time-varying 6DoF transformations are predicted by a deformation MLP and propagated to canonical Gaussians through Linear Blend Skinning (LBS). Similarly, 4DGS-SLAM incorporates motion control deformation into RGB-D SLAM, optimizing dynamic motion through learned control-point transformations and optical-flow supervision.

In contrast, TAGs explicitly model dynamic motion using long-lived 3D trajectories recovered from 2D point tracking, predicted depth, and optimized camera poses, without requiring an additional learned deformation network. More importantly, TAGs maintain persistent Gaussian-to-track associations across observations, enabling canonical geometry and appearance to be consistently accumulated over time. Combined with coverage-aware Gaussian insertion, this design incrementally integrates newly observed dynamic content into a persistent 4D map without repeated cross-frame Gaussian matching.

To further validate the effectiveness of our anchoring design, we conduct an additional ablation study by replacing TAGs with a dynamic deformation model based on control points and an MLP. As shown in Table R2-1, TAGs improve PSNR by 1.45 dB and SSIM by 0.021, and reduce LPIPS by 0.033, demonstrating the effectiveness of our track-anchored dynamic representation.

Table R2-1. Ablation study comparing TAGs with control-point-based MLP deformation.

PSNRSSIMLPIPS
w/ Control Points + MLP25.830.8810.189
w/ TAGs27.280.9020.156

Q2. Track loss and reappearance

A2: Track reliability is determined by two criteria: (1) the visibility probability (\sigma(v)) predicted by Track-On-R must be at least 0.8, and (2) the tracked pixel must lie within the predicted dynamic region.

When a track loses reliable support, its associated Gaussians are temporarily hidden and excluded from rendering, while their canonical parameters and associations are preserved. Once the object reappears, if Track-On-R successfully recovers the original track identity, the corresponding Gaussians can be reactivated for rendering with their existing associations. Otherwise, new tracks and Gaussians are initialized.

Our current object tracking relies on Track-On-R. When an object completely leaves the camera's field of view and subsequently re-enters, Track-On-R may fail to recover its historical trajectory and instead initialize a new track. We acknowledge this limitation and plan to incorporate long-term object re-identification mechanisms in future work.

AuthorsOpenReview ↗

Respomse to Reviewer PBRm (2/2)

Q3. Propagation of tracking and depth errors

A3: To mitigate the impact of inaccurate tracking and depth estimation, we discard trajectories and Gaussians that move outside the predicted dynamic mask regions. Furthermore, when camera poses are updated, the corresponding 3D track positions and Gaussian centers are adjusted accordingly. Since both are reconstructed from the same image-space observations using depth estimates and camera poses, their relative spatial relationships are preserved, and the established Gaussian-to-track associations do not require reassignment. We further conduct an ablation study to evaluate the effectiveness of this filtering strategy. As shown in Table R2-2, disabling the filtering of out-of-mask Gaussians results in a slight decrease in rendering PSNR.

Table R2-2. Ablation study on filtering Gaussians outside dynamic mask regions.

MethodPSNR ↑SSIM ↑LPIPS ↓
w/o Out-of-Mask Filtering27.220.9020.156
Full Model (Ours)27.280.9020.156

Q4&W2. Evidence for occlusion robustness and temporal persistence

A4: Since the Bonn dataset does not provide ground-truth trajectories for dynamic objects, we conduct an additional quantitative analysis of track persistence and recovery under occlusion. Specifically, we report the number of dynamic trajectories, average track survival rate, number of tracks recovered after occlusion, and average occlusion duration for each sequence. As shown in Table R2-3, our method achieves an average track survival rate of at least 85% across all evaluated sequences. These results provide additional quantitative evidence that our approach can maintain long-lived motion trajectories and recover their original associations following temporary occlusion, supporting the effectiveness of our persistent track-anchored representation.

Table R2-3. Quantitative evaluation of track persistence and recovery under occlusion on the Bonn dataset.

SceneBalloonBalloon2CrowdCrowd2MovingMoving2PersonPerson2
Number of Tracks3756704731063335474298307720021931
Average Survival Rate87.3%85.9%91.2%91.7%90.2%90.5%87.3%85.0%
Tracks Recovered After Occlusion1379258066805973458174582695
Mean Occlusion Duration (frames)26.329.511.715.822.517.919.018.3
Reviewer_YnDNOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Monocular dynamic Gaussian SLAM on a WorldMirror backbone with one added dynamic head. CAPS computes Shannon entropy of renormalised cross-view attention at block 14, heads {1,4,10}, and biases which of 128 patches enters bundle adjustment. Static mapping voxelizes prior Gaussians and inserts them only where rendered opacity < 0.9. Dynamic content uses canonical Gaussians bound at birth to the nearest 3D motion track, whose displacement drives the centre, with 4D spherical harmonics for appearance. Tested on Bonn and Wild-SLAM MoCap; the segmentation head on DAVIS.

Critical strengths
  • The decomposition is right and the ablation proves it load-bearing. Removing TAGs costs 6.8 dB on Bonn (27.28 → 20.47), removing coverage-aware insertion 5.0 dB, removing 4DSH 0.29. Binding content to persistent tracks instead of re-matching Gaussians per frame attacks the failure mode this literature keeps hitting.
  • Segmentation is strong and honestly validated. 80.1/91.8 JM/JR on DAVIS-16 against 68.1/78.3 for π3mos-SLAM, from a head trained on eight datasets none of which is DAVIS.
  • Fast for what it does: 0.94 FPS against 0.36 for Flow4DGS-SLAM and 0.04 for 4DGS-SLAM, mostly from cheaper mapping.
Critical weaknesses
  • Motion is translation only. Eqs. (8)–(9) add the anchor's displacement to a frozen canonical centre: no rotation, no scale change, and the binding is fixed at birth with no re-binding when the nearest track changes. A rotating or deforming object is not representable, and these are what Wild-SLAM is made of. Since each Gaussian binds independently, neighbours on one rigid object can bind to different tracks and tear. Claims of robustness under occlusion oversell what the representation expresses.
  • The tracking lead is not established. Bonn average is 2.22 cm against 2.30 (DROID-W) and 2.37 (π3mos-SLAM), with losses on Balloon (2.7 versus 2.5), Balloon2 (2.5 versus 2.4) and Crowd2 (2.6 versus 1.8). On Wild-SLAM the 0.56 cm average hides losses on Person (0.6 versus 0.2) and Table1 (0.7 versus 0.4); the win comes entirely from Table2, where baselines score 2.1–2.7 and this method 0.9. No run-to-run deviation for either table.
  • Rendering comparisons are partly protocol. Table 5 recomputes the two RGB-D baselines on full images where the originals masked to valid depth, and includes no monocular dynamic method; those baselines also use roughly 25× the loss evaluations.
  • Nothing tests persistence. Both datasets are short, object-centric clips; no sequence where an object leaves and returns, no multi-minute run, no count of surviving tracks — yet persistence is the central claim of a 4D map.
Questions for authors
  1. How often does the nearest track exceed 0.02, and what happens to unbound Gaussians?
  2. Can a Gaussian be re-bound later, and what stops tearing inside one object?
  3. Report deviations for Tables 2 and 3 and explain the Person and Table1 losses.
4 comments
AuthorsOpenReview ↗

Respomse to Reviewer YnDN (1/3)

We sincerely thank the reviewer for the constructive and technically insightful feedback, as well as for recognizing the effectiveness of our track-anchored representation, dynamic segmentation, and computational efficiency. We appreciate the concerns regarding motion expressiveness, tracking reliability, evaluation fairness, and temporal persistence. Our point-by-point responses are provided below.

W1. Motion expressiveness

A1: We acknowledge that TAGs do not explicitly parameterize object-level SE(3) transformations, as object motion is modeled through trajectory anchors. However, we would like to clarify that although each Gaussian is propagated through the translation of its associated anchor, the dynamic object as a whole is not restricted to a single global translation. Specifically, different Gaussians can be associated with different 3D motion trajectories, allowing their centers to undergo spatially varying displacements. Consequently, the collective motion of Gaussian centers can approximate non-rigid object deformation to a certain extent. Figure 3 and the supplementary videos demonstrate our method's capability to reconstruct non-rigid human motion.

Furthermore, each Gaussian is associated with its nearest trajectory anchor, and Gaussians sharing the same anchor undergo identical displacements, ensuring coherent motion within each anchor group without tearing. However, when neighboring Gaussians are associated with different trajectories, tracking errors may introduce inconsistent motion and cause local geometric tearing. Incorporating object-level rigidity constraints or local deformation consistency regularization is a promising direction for future improvement.

Regarding occlusion, our method preserves canonical Gaussian parameters when their associated tracks are temporarily occluded. Once the object reappears and the corresponding track provides valid support again, these Gaussians can be reactivated for rendering.

W2. Marginal tracking improvements and statistical reliability

A2: Although rankings vary across sequences, Tables 2 and 3 show that TAG4D-SLAM attains the lowest average ATE on both benchmarks. Table R1-1 reports the mean and standard deviation of this dataset-average ATE over five runs for DROID-W, π³mos-SLAM, and our method (cm). Our means are 2.22 cm on Bonn and 0.56 cm on Wild-SLAM MoCap, with standard deviations of 0.01747 cm and 0.00266 cm. The multi-run standard deviation of π³mos-SLAM is an order of magnitude higher than ours, 0.186 versus 0.01747 on Bonn (about 11×) and 0.1068 versus 0.00266 on Wild-SLAM MoCap (about 40×). Taken together, our method is stably better than the other two baselines.

This stable advantage comes from our tracking contribution, CAPS. It converts the feed-forward model’s internal cross-view attention into a sampling prior for bundle adjustment. The ablation in Table 7 demonstrates its effectiveness within the same tracker: replacing uniform sampling with CAPS reduces average Bonn ATE by 14%.

Table R1-1. Mean and standard deviation of dataset-average ATE over five runs (cm)

DROID-Wπ³mos-SLAMOurs
Bonn Avg./Std.2.30/0.009242.37/0.1862.22/0.01747
Wild-SLAM MoCap Avg./Std.0.96/0.004310.66/0.10680.56/0.00266
AuthorsOpenReview ↗

Respomse to Reviewer YnDN (2/3)

W3. Rendering Evaluation Protocol

A3: We thank the reviewer for highlighting the importance of evaluation fairness. Both 4DGS-SLAM and Flow4DGS-SLAM were evaluated strictly following their official implementations. The only modification was in the computation of PSNR for rendered RGB images, where we evaluated the entire image instead of using the valid-depth-masked regions adopted in their original papers. This ensures that all methods are evaluated over the same full-image domain. For LPIPS and SSIM, both 4DGS-SLAM and Flow4DGS-SLAM already employ full-image evaluation in their original papers; therefore, our reported results are consistent with their original evaluations.

Regarding optimization budgets, our method performs 80 mapping iterations per keyframe, whereas 4DGS-SLAM and Flow4DGS-SLAM perform 200 iterations per keyframe, with losses accumulated over 10 frames in each iteration. Consequently, under the reported configurations, these baselines require approximately 25 times more loss evaluations than our method.

Despite using less input information (monocular RGB without ground-truth depth) and a substantially smaller optimization budget, our method achieves superior tracking and mapping performance compared with 4DGS-SLAM and Flow4DGS-SLAM. Furthermore, our system achieves 0.94 FPS, substantially outperforming Flow4DGS-SLAM (0.36 FPS) and 4DGS-SLAM (0.04 FPS), demonstrating its computational efficiency.

Regarding the absence of a monocular dynamic reconstruction baseline, we have added a comparison with De4D-SLAM (Fan et al., ICML 2026), which also uses RGB-only input. The De4D-SLAM results in Table R1-2 are quoted from its original paper. Across 12 Bonn sequences, our method achieves an average PSNR of 27.31 dB, compared with 26.70 dB for De4D-SLAM, an improvement of 0.61 dB.

Table R1-2. Rendering comparison with De4D-SLAM on 12 Bonn sequences. De4D-SLAM results are quoted from its original paper.

MethodBalloonBal.2PersonPer.2CrowdCrwd.2SyncSync2Rem.Rem.2PlacePlace2PSNR↑SSIM↑LPIPS↓
De4D-SLAM [RGB]27.3128.4327.4927.1527.4928.7827.8823.2828.6622.7925.2425.8926.700.9030.200
Ours [RGB]28.7027.1827.8626.9826.6226.3929.4630.0026.0825.3926.3626.7227.310.9060.150

W4. Evidence for Persistence.

A4: Bonn and Wild-SLAM are widely adopted benchmarks for dynamic SLAM evaluation. Their sequences range from several hundred to over one thousand frames. In particular, the Crowd and Crowd2 sequences contain multiple people repeatedly entering and leaving the camera's field of view. The supplementary videos visualize our method's ability to continuously track and reconstruct dynamic objects throughout these sequences.

Specifically, we report the number of dynamic trajectories, average track survival rate, number of tracks recovered after occlusion, and average occlusion duration for each sequence. As shown in Table R1-3, our method achieves an average track survival rate of at least 85% across all evaluated sequences. These results provide additional quantitative evidence that our approach can recover associations following temporary occlusion.

Table R1-3. Quantitative evaluation of track persistence and recovery under occlusion on the Bonn dataset.

SceneBalloonBalloon2CrowdCrowd2MovingMoving2PersonPerson2
Number of Tracks3756704731063335474298307720021931
Average Survival Rate87.3%85.9%91.2%91.7%90.2%90.5%87.3%85.0%
Tracks Recovered After Occlusion1379258066805973458174582695
Mean Occlusion Duration (frames)26.329.511.715.822.517.919.018.3

Our current method can continuously track and reconstruct objects as long as their motion trajectories remain available. However, when an object completely leaves the field of view and subsequently reappears, the system may fail to recover its previous identity and instead initialize it as a new object. We acknowledge this limitation and plan to improve long-term object re-identification and association in future work.

AuthorsOpenReview ↗

Respomse to Reviewer YnDN (3/3)

Q1. How often does the nearest track exceed 0.02, and what happens to unbound Gaussians?

A5: We have additionally measured the proportion of dynamic Gaussian candidates whose distance to the nearest valid track exceeds the threshold of 0.02, which is only 4.29%. These candidates are primarily outliers caused by inaccurate depth estimation, particularly around object boundaries where depth artifacts and trailing effects may occur. Such Gaussians are directly discarded to prevent erroneous geometric content from interfering with dynamic object reconstruction.

Q2. Can a Gaussian be re-bound later, and what stops tearing inside one object?

A6: In our current implementation, Gaussian-to-track associations are established at initialization and remain fixed throughout subsequent dynamic map optimization. Gaussians sharing the same trajectory anchor undergo identical center displacements, ensuring locally coherent motion within each anchor group. However, no explicit common rigid transformation constraint is imposed across different anchors. Consequently, significant tracking errors or inconsistent motion between neighboring trajectories may lead to local geometric discontinuities or tearing.

To mitigate such artifacts, our opacity-guided Gaussian insertion strategy introduces new Gaussians into under-reconstructed regions, including gaps caused by inconsistent motion, and associates these newly inserted Gaussians with valid trajectory anchors. We acknowledge that incorporating adaptive Gaussian-to-track re-association or explicit local rigidity constraints would be a valuable direction for future improvement.

Q3. Report deviations for Tables 2 and 3 and explain the Person and Table1 losses.

A7: Table R1-1 reports the five-run mean and standard deviation of the dataset-average ATE. For Person and Table1, relative to the distance traveled, these errors are small. Over trajectories of 10.35 m on Person and 6.59 m on Table1, our ATE corresponds to only 0.053% and 0.104% of the respective trajectory lengths. The gaps to π³mos-SLAM amount to just 0.032% and 0.028% of those lengths, further illustrating the high absolute tracking accuracy achieved on both sequences. Our analysis of relative pose evaluation, local similarity alignment, and the fitted local scales shows that gradual scale drift is the main explanation for these minor gaps. In future work, we will explore stronger global constraints.

Dear Reviewer YnDN,

Thank you again for taking the time to review our paper and share your comments. We hope our earlier response has addressed your concerns. Since the ICLR discussion period ends soon, we would greatly appreciate your feedback on our response, and we are happy to answer any further questions.

Best regards,

The Authors

Comments on the paper

2 comments

Thank you for the positive feedback and for pointing out these closely related formulations. We agree that canonical-motion decoupling, anchor-based representations, and adaptive densification have been explored in previous works. We further clarify the technical distinctions and contributions of TAGs below.

1. Observation-driven Gaussian-to-track association. SC-GS and 4DGS-SLAM associate Gaussians with neighboring learnable control points and interpolate their MLP-predicted transformations to construct dense deformation fields. Anchored 4DGS instead uses learnable spatiotemporal anchors to generate local 4D Gaussian attributes. MoSca is more closely related to TAGs, as it also lifts 2D tracks into 3D. However, it uses these tracks to initialize a globally optimized SE(3) deformation graph, with multi-node blending and learnable skinning weight corrections.

In contrast, TAGs directly associate each newly observed Gaussian with a valid 3D trajectory at its birth time and compensate for the anchor's displacement to initialize its canonical position. The trajectory identity and association remain fixed during map refinement. This provides a persistent motion reference for Gaussians introduced at different timestamps, without optimizing an additional deformation field or skinning weights. Furthermore, separating track-conditioned visibility from canonical Gaussian parameters allows previously reconstructed content to be retained during temporary occlusion and reused when the original track recovers.

2. Coupled expansion of motion anchors and dynamic content. Unlike Anchored 4DGS, which grows anchors according to accumulated gradients, motion velocities, and temporal durations, TAGs expand the dynamic representation according to two complementary observation-driven criteria: (i) new tracking queries are introduced in dynamic regions insufficiently covered by existing trajectories; and (ii) existing Gaussians are propagated to the current timestamp, and new feed-forward Gaussians are inserted only where the rendered dynamic opacity indicates insufficient coverage. Each inserted Gaussian is immediately associated with a valid trajectory and transformed into its canonical representation.

This design distinguishes newly required dynamic content from content already represented by the moving map, enabling selective integration of newly visible surfaces while limiting redundant insertion. Notably, 4DGS-SLAM does not insert new dynamic Gaussians after keyframe insertion, while MoSca initializes and globally fuses Gaussians across the input video. TAGs instead progressively expand both motion support and scene content as new observations arrive, without requiring future observations.

3. Experimental evidence. The additional comparison in Table R2-1 shows that replacing the control-point-based MLP deformation with TAGs improves PSNR from 25.83 to 27.28 dB. Disabling coverage-aware updates in Table 7 reduces PSNR to 22.24 dB, demonstrating the importance of dynamic expansion.

Dear Reviewer PBRm,

Thank you again for your valuable and constructive feedback. As the discussion period is approaching its deadline, we would appreciate knowing whether our responses have adequately addressed your concerns. If any questions remain, we would be happy to provide further clarification.

We sincerely appreciate your time and consideration and look forward to your feedback.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.