acceptodds
Under review as a conference paper at ICLR 2027

TrajForward: Feed-Forward Dynamic 3D Gaussian Reconstruction for Urban Driving Scenes with Explicit Motion Trajectories

Abstract

Reconstructing dynamic driving scenes with feed-forward 3D Gaussian Splatting remains fundamentally challenged by temporal inaccuracies and under-constrained entity-level correspondence, leading to limited dynamic expressiveness, unstable aggregation, and novel-view artifacts. These issues stem from the local or discrete low-order approximation and the lack of explicit entity-level constraint under self-supervision. To address this, we propose to derive target-time Gaussian centers in an entity-aware manner that not only captures coherent nonlinear motions but also preserves persistent temporal correspondence and intra-entity local variations. We materialize this concept through TrajForward, a feed-forward dynamic 3DGS framework that instantiates Gaussian centers along entity-conditioned trajectories. Specifically, to resolve entity-association ambiguity, we introduce learnable motion-bearing slots as latent proxies. These slots bind dynamically related Gaussians, providing explicit structural coupling and stable cross-time correspondence without object annotations. Building upon these proxies, we tackle heterogeneous nonlinear dynamics through a hierarchical trajectory representation. By decomposing motion into slot-level macroscopic trajectories and Gaussian-specific local residuals, this elegantly preserves both entity coherence and local flexibility. Finally, to prevent self-supervised trajectory instability, we formulate a suite of trajectory-aware objectives. By explicitly regularizing slot assignments and trajectory smoothness, these objectives mitigate static-dynamic entanglement and promote temporally coherent, artifact-reduced reconstructions. Experiments on challenging urban driving benchmarks validate TrajForward and demonstrate improved rendering fidelity. Code will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Feed-Forward Reconstruction, Dynamic 3D Gaussian Splatting, Autonomous Driving, Image-Based Rendering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JQ23OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TrajForward, a feed-forward framework for dynamic driving-scene reconstruction using 3D Gaussian Splatting. TrajForward tackle the dynamic problem by 3 core contribution:

  1. Motion-bearing slots: Learned slots group Gaussians through hard assignments, encouraging shared motion within each group without instance annotations.
  2. Hierarchical continuous trajectories: Shared slot-level trajectories capture coarse motion, while Gaussian-specific residual trajectories accommodate local variations.
  3. Trajectory-aware training objectives: Rendering supervision is complemented by regularization of trajectory smoothness, slot assignments, control-point magnitudes, and static-region motion. Experiments mainly evaluate rendering quality and motion estimation on Waymo, with additional cross-dataset results on nuScenes and Argoverse 2.
Critical strengths

The method uses slots to combine shared motion with local residuals for dynamic objects. And Tab. 1–4 show improvements in rendering quality and motion accuracy, supporting the effectiveness of the method.

Critical weaknesses

None

Questions for authors

Main concerns: 1.Baseline comparisons. Could the authors explain why SLARM and UniSplat are not included in the experiments? Comparisons under matched settings would help assess the improvements over recent methods. 2.The full model achieves different dynamic PSNR values in Tables 1, 8, and 9 (23.11, 23.78, and 23.74). Are these results obtained under different settings? Also, Table 2 describes Waymo-to-Argoverse 2 transfer, while Appendix A describes nuScenes-to-Argoverse 2 transfer. Which training dataset is used?

Minor concern: In Table 8, are the polynomial, Bézier, and B-spline trajectories compared with similar numbers of parameters? This would help explain the performance differences

5 comments
AuthorsOpenReview ↗

Response to Reviewer JQ23 (Part 1/3): Question 1

Question 1. Baseline comparisons. Could the authors explain why SLARM and UniSplat are not included in the experiments? Comparisons under matched settings would help assess the improvements over recent methods.

We thank the reviewer for the valuable suggestion. However, UniSplat and SLARM do not fully follow the evaluation protocol of STORM/DGGT. On Waymo, STORM/DGGT divide each sequence into non-overlapping 2-second clips of 20 frames, using the 1st, 5th, 10th, and 15th frames as context and evaluating NVS on the remaining 16 frames. In contrast, UniSplat primarily synthesizes novel views at the subsequent timestep (t+1) from images at (t). Although SLARM adopts a similar sparse-input setting, its paper does not explicitly specify the target frames, and its released evaluation code defaults to evaluating only the 12 interpolated frames, excluding the final four extrapolated frames. Thus, directly comparing their reported results would not ensure a fair evaluation, which is why we did not include them in our main experiments.

Following the reviewer's suggestion, we adapted UniSplat to the unified STORM/DGGT training and evaluation protocol, as shown in the below table. Its performance degradation may be attributed to UniSplat's reliance on ego-motion compensation and implicit temporal feature fusion, without explicit motion alignment for independently moving objects, making it difficult to handle large dynamic displacements under sparse observations.

Regarding SLARM, although it adopts a third-order Taylor expansion to improve nonlinear motion modeling, it essentially still follows the increment-based propagation paradigm described in our Introduction. Since SLARM independently predicts motion coefficients for each Gaussian without explicit entity-level shared motion constraints, it is difficult to ensure coherent motion of entire dynamic objects and persistent cross-time correspondence. Nevertheless, SLARM introduces multiple forms of semantic supervision to indirectly constrain object-level motion coherence and temporal correspondence. In contrast, our method explicitly models motion relationships among related Gaussians through a hierarchical formulation of shared continuous trajectories and local residuals, enhancing entity-level temporal consistency while preserving local motion flexibility.

We also re-evaluated SLARM under the unified evaluation protocol. Although the full SLARM model achieves dynamic rendering performance comparable to ours, it additionally relies on LSeg semantic feature distillation and GT semantic label supervision. Our ablation results show that removing GT semantic label supervision leads to only a minor drop in full-image PSNR (0.17 dB), but a substantially larger decrease in dynamic-region PSNR (0.55 dB). This suggests that semantic supervision primarily benefits dynamic reconstruction and may partially compensate for the lack of explicit entity-level shared motion constraints in SLARM. In contrast, our method achieves superior dynamic reconstruction quality (23.11 vs. 22.90 dB) with substantially faster inference (0.12 vs. 0.54 s), without additional semantic supervision, demonstrating the effectiveness of structured shared trajectories and local residual modeling.

We will expand the Related Work section to clarify the methodological differences between TrajForward, UniSplat, and SLARM, and update the experimental section with their results under the unified evaluation protocol to provide a more comprehensive and fair comparison.

Table: Full-sequence and future-frame rendering comparisons

MethodFull sequence Dynamic-only PSNR↑Full sequence Dynamic-only SSIM↑Full sequence Full image PSNR↑Full sequence Full image SSIM↑Future Next 1st Dynamic-only PSNR↑Future Next 2nd Dynamic-only PSNR↑Future Next 3rd Dynamic-only PSNR↑Future Next 4th Dynamic-only PSNR↑Inference speed
STORM22.100.62426.380.79421.7921.5421.5221.130.18s
UniSplat21.050.60226.500.79119.8519.4719.1218.613.37s
SLARM-F22.900.63327.190.82022.7622.5821.9921.760.54s
SLARM-F w/o segmentation labels22.350.61627.020.80922.1321.7721.0320.580.55s
Ours23.110.64127.130.82923.1023.0822.9822.560.12s

Table: Scene Flow Estimation Comparison

MethodEPE3DEPE_{3D} (m) ↓Acc5Acc_5 (%) ↑Acc10Acc_{10} (%) ↑θ\theta (rad) ↓
NSFP0.69842.1754.260.919
NSFP++0.71153.1063.020.989
STORM0.27681.1285.610.658
UniSplat0.37564.3473.980.787
SLARM0.24377.4983.910.646
SLARM w/o segmentation labels0.29273.1480.030.665
Ours0.17988.3892.400.317
AuthorsOpenReview ↗

Response to Reviewer JQ23 (Part 2/3): Question 2

Question 2. The full model achieves different dynamic PSNR values in Tables 1, 8, and 9 (23.11, 23.78, and 23.74). Are these results obtained under different settings? Also, Table 2 describes Waymo-to-Argoverse 2 transfer, while Appendix A describes nuScenes-to-Argoverse 2 transfer. Which training dataset is used?

(1) The discrepancy primarily arises from different camera configurations. Table 1 uses three frontal cameras, whereas Tables 8 and 9 adopt a single-camera setting to reduce computational costs and facilitate extensive ablation studies, as the factors under investigation are not directly related to the number of cameras. This difference in evaluation settings accounts for the higher PSNR values reported in Tables 8 and 9. The minor discrepancy between these two tables (23.78 vs. 23.74 dB) results from separate experimental runs conducted at different times and reflects small run-to-run variations. We will explicitly clarify these configuration differences in the revised manuscript to avoid ambiguity.

(2) We apologize for the ambiguity caused by a typo in Appendix A. Consistent with Table 2, our generalization experiments involve two distinct settings: (1) in-domain training and evaluation on nuScenes, and (2) zero-shot evaluation on Argoverse 2 using the Waymo-trained model, without additional fine-tuning. To avoid confusion, we will correct the statement in Appendix A from "In this setting, models are trained and evaluated on nuScenes, and then directly evaluated on Argoverse 2 without additional fine-tuning." to "In this setting, models are trained and evaluated on nuScenes, while the Waymo-trained models are directly evaluated on Argoverse 2 without additional fine-tuning."

AuthorsOpenReview ↗

Response to Reviewer JQ23 (Part 3/3): Question 3

Question 3. In Table 8, are the polynomial, Bézier, and B-spline trajectories compared with similar numbers of parameters? This would help explain the performance differences

Thank you for this valuable suggestion. Specifically, the linear (degree-1), quadratic (degree-2) and cubic (degree-3) polynomials employ 2, 3 and 4 learnable 3D coefficient vectors, corresponding to 6, 9 and 12 scalar parameters, respectively. The cubic Bézier (degree-3) and quadratic B-spline (degree-2) each employ four learnable 3D control points (12 scalar parameters), with fixed knots for the B-spline. Under matched parameter budgets, Bézier (23.78 dB) and B-spline (23.44 dB) outperform the cubic polynomial (22.38 dB), indicating that the performance differences cannot be solely attributed to parameter count, but also to trajectory parameterization and optimization characteristics. We will explicitly clarify these configurations in the revised appendix.

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

TrajForward reconstructs dynamic driving scenes as 3D Gaussians in a single forward pass from posed images at several timestamps. Its central idea is to share motion through learned slots while retaining Gaussian-specific flexibility. A Transformer predicts Gaussian anchors, slot assignments and continuous trajectory parameters. Each Gaussian follows one slot in the forward pass; its center combines the slot's displacement with a residual trajectory gated by its dynamic-assignment probability (Fig. 2; Eqs. 9–16). Rendering supervision and trajectory regularizers train this decomposition without object or motion annotations.

The experiments evaluate dynamic-region rendering, future frames, motion estimation and slot correspondence, alongside non-uniform-motion and longer-clip settings. The main contribution is a feed-forward implementation of structured continuous motion, with empirical evidence that it helps reconstruct moving objects from sparse observations.

Critical strengths
  1. The representation offers a sensible balance between shared motion and local variation. Figure 2 and Eqs. 13–16 give Gaussians a common continuous path without forcing every point to move identically. The residual trajectory preserves local flexibility, while the static slot and gating discourage background motion. This is a useful inductive bias for sparse driving observations.

  2. The strongest results address the intended motion problem. In Table 1, TrajForward maintains a dynamic-region advantage over STORM across all four future-frame columns, including 22.56 versus 21.13 dB at the fourth target frame. Tables 6–7 also report favorable results on non-uniform motion and four-second clips. I do appreciate Table 4.

  3. The paper examines what the latent slots represent. Table 5 and Fig. 5 assess correspondence rather than inferring object structure from rendering quality alone.

Critical weaknesses
  1. The comparison provenance is not clear enough to establish how much of the reported advantage is measured under matched conditions. Section 4.1 (p.7, L363–372) specifies context frames and metrics, but does not identify which baseline entries in Tables 1–4 were reproduced or taken from prior reports, with which checkpoints and evaluation resolution. The same issue affects the speed comparison in Sec. 4.2 (p.8, L408–409), where hardware and timed operations matter.

  2. Concern abouth the novelty and scope. The isolated-Gaussian contrast in Fig. 1 and p.2, L79–81 is too broad for STORM, which already combines shared velocity bases through learned motion tokens; the submission itself acknowledges this in Sec. 2, L145–146. BézierGS also models an object's center and Gaussian offsets with separate Bézier curves, although in a different reconstruction setting.

  3. Table 1 reports dynamic PSNR/SSIM of 23.11/0.641, while the full objective in Table 9 and K=4 in Table 10 give 23.74/0.648. Section 4.3 (p.9, L458–475) does not explain this difference.

Questions for authors
  1. Could you clarify the provenance and shared evaluation conditions of the central rendering and flow comparisons, including which trained model produced the Argoverse 2 results?

  2. How do the headline model in Table 1 and the defaults in Tables 8–10 correspond? Could you explain the configuration differences, including the curve degree and metric region where these are currently unspecified.

  3. About the contribution scope. Is it fair to understand the main advance as annotation-free, feed-forward learning of shared continuous trajectories and local residuals, with slots representing coherent motion groups?

11 comments
AuthorsOpenReview ↗

Response to Reviewer dCFq (Part 1/6): Weakness 1

Weakness 1. The comparison provenance is not clear enough to establish how much of the reported advantage is measured under matched conditions. Section 4.1 (p.7, L363–372) specifies context frames and metrics, but does not identify which baseline entries in Tables 1–4 were reproduced or taken from prior reports, with which checkpoints and evaluation resolution. The same issue affects the speed comparison in Sec. 4.2 (p.8, L408–409), where hardware and timed operations matter.

Thanks for your advice. We clarify the details as follows.

(1) Baseline provenance. We will explicitly distinguish results taken from prior publications (marked with †) from those obtained through our own reproduction in Tables 1–4.

  • Table 1 (Waymo): For full-sequence evaluation, we use the published results of STORM, DGGT, and PVG (with PVG results reported in STORM/DGGT). We also independently reproduced these baselines and obtained comparable results, with PSNR differences of only 0.2–0.4 dB below the reported values. Therefore, we retain the published numbers. Since none of the baselines originally reported future-frame evaluation, we evaluate STORM, DGGT using their publicly released GitHub checkpoints. For the remaining baselines (ReconDrive, DeSiRe-GS, PVG, AnySplat, DA3, DrivingForward, and DepthSplat), we conduct our own reproductions, as their original evaluation protocols and scenes differ from those used by STORM/DGGT.

  • Table 2 (Cross-dataset generalization): All results are from our own experiments. Unlike STORM, which evaluates models trained and tested separately on nuScenes and Argoverse 2, we have improved it by adopting two settings: (i) in-domain training and evaluation on nuScenes, and (ii) zero-shot transfer to Argoverse 2 using Waymo-trained models. Furthermore, STORM and DGGT report only full-image metrics, without providing dynamic-region metrics. We therefore conduct our own evaluations to ensure a consistent and comprehensive comparison across both settings.

  • Table 3: All baseline results are obtained from our own reproductions under the evaluation protocol described in the paper.

  • Table 4: Baseline results are taken from STORM and DGGT, except for FastNSF, which we additionally evaluate. The literature-reported entries will be explicitly marked with †.

(2) Evaluation resolution. Following STORM, our evaluations use a resolution of 160 × 240 on Waymo, 160 × 288 on nuScenes, and 192 × 256 on Argoverse 2. These resolutions are consistently applied to our reproduced baseline evaluations.

(3) Hardware and inference speed. Following STORM and DGGT, our inference-speed measurements are conducted on a single NVIDIA A100 GPU, with runtime covering both feed-forward network inference and Gaussian generation.

We will incorporate these clarifications into the revised manuscript, including explicit provenance annotations, evaluation resolutions, checkpoint sources, and inference-timing details, to make the comparison protocol and reproducibility more transparent.

AuthorsOpenReview ↗

Response to Reviewer dCFq (Part 2/6): Weakness 2

Weakness 2. Concern abouth the novelty and scope. The isolated-Gaussian contrast in Fig. 1 and p.2, L79–81 is too broad for STORM, which already combines shared velocity bases through learned motion tokens; the submission itself acknowledges this in Sec. 2, L145–146. BézierGS also models an object's center and Gaussian offsets with separate Bézier curves, although in a different reconstruction setting.

Thank you for this valuable suggestion. We would like to further clarify the key differences between our TrajForward, STORM, and BézierGS.

(1) Distinction from STORM: From shared motion bases to explicit entity-level structural associations.
We first clarify that the term isolated-Gaussian in our paper does not imply that Gaussian motions are predicted entirely independently, but rather refers to the lack of explicit structural associations between Gaussians and potential dynamic entities. Specifically, STORM predicts a set of shared abstract velocity bases and combines them using Gaussian-specific weights to obtain individual Gaussian velocities. Although this introduces shared motion information, it does not explicitly establish structural associations between Gaussians and coherent motion entities. Consequently, motion is still represented primarily at the individual Gaussian level. In contrast, TrajForward explicitly associates Gaussians with latent motion entities through learnable motion-bearing slots, enabling coherently moving Gaussians to share continuous trajectories while preserving local motion variations through Gaussian-specific residuals. Therefore, our key distinction is not whether shared motion information is utilized, but whether explicit entity-level motion associations are established and leveraged for hierarchical continuous trajectory modeling.

(2) Distinction from BézierGS: Annotation-free feed-forward reconstruction and generalizable trajectory modeling.
Although BézierGS also models object-center motion and Gaussian-level offsets using Bézier curves, the two approaches differ fundamentally in their reconstruction paradigms and trajectory formulations. First, BézierGS relies on 3D bounding boxes to separate dynamic objects and performs scene-specific optimization for individual objects. In contrast, TrajForward does not require 3D bounding boxes or explicit object annotations. Instead, it automatically learns motion associations through latent slots and reconstructs dynamic scenes in a single feed-forward pass. Second, while BézierGS is specifically designed around Bézier-based motion modeling, TrajForward provides a more general hierarchical continuous trajectory framework that is not restricted to a particular trajectory parameterization. It supports linear, polynomial, B-spline, and Bézier trajectories, as demonstrated in Table 8 of the main paper. On the overall Waymo benchmark, Bézier achieves the best performance, while B-spline shows certain advantages in scenes with more pedestrians and complex non-rigid motion, as demonstrated in Table 12 of the Appendix. These observations suggest that different trajectory parameterizations may offer complementary strengths. We have discussed these findings in Section 4.3 in the main paper and Appendix D, E and reiterate them here.

Table 8: Comparison of different trajectory parameterizations. (from our main paper)

MethodPSNR↑SSIM↑
Linear21.030.598
Quadratic22.010.617
Cubic22.380.622
B-Spline23.440.631
Bézier23.780.649

Table 12: Comparison between Bézier and B-spline trajectory parameterizations on challenging motion subsets. (from our main paper)

Evaluation SubsetBézierB-spline
Car-dense clips21.8621.54
High-density non-rigid-actor clips20.3520.71
AuthorsOpenReview ↗

Response to Reviewer dCFq (Part 3/6): Weakness 3

Weakness 3. Table 1 reports dynamic PSNR/SSIM of 23.11/0.641, while the full objective in Table 9 and K=4 in Table 10 give 23.74/0.648. Section 4.3 (p.9, L458–475) does not explain this difference.

Our main experiments (Table 1) adopt a three-camera setting, yielding a dynamic-region PSNR of 23.11 dB. Since the ablation studies primarily investigate the effects of different components and parameter configurations, which are not directly related to the number of cameras, we adopt a single-camera setting for Tables 9 and 10 (PSNR 23.74dB) to reduce computational costs and facilitate extensive controlled experiments. We will explicitly clarify these configuration and evaluation differences in the revised manuscript to avoid ambiguity and improve reproducibility.

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

The paper proposes a new feed-forward method for dynamic 3D reconstruction in urban driving scenarios. A key ingredient for better motion estimation seems the hierarchical modeling of motion trajectories in combination with a slot assignment for each moving object instance.

Critical strengths

The method is well motivated and especially the proposed hierarchical trajectory modeling across longer time periods for individual object instances seems a plausible approach to mitigate the fitting to local minima which is a common problem for models that only estimate the motion field between consecutive frames. The proposed method achieves competitive reconstruction results.

Critical weaknesses
  1. For a method that reconstructs dynamic scenes, I would have liked to see a re-rendered video of the reconstruction results in the supplementary material in order to be able judge the quality of the reconstruction. Furthermore, a visualization of the estimated motion fields would be equally useful.

  2. The discussion of method limitations would be helpful.

Minor: The term "4D Gaussian splatting" is a misnomer since temporally varying 3D Gaussians are not 4D Gaussians (Gaussians are well-defined in any dimension with an n-dimensional mean vector and n x n dimensional covariance matrix). The community should do an effort to avoid to use of incorrect/confusing terminology and not further propagate such naming mistakes. There are better alternatives, e.g. "dynamic Gaussian splatting".

Overall, the method is well-motivated and seems technically sound. The design decisions are well justified and empirically supported in the experimental evaluation which yields state-of-the-art results. Therefore, I am leaning towards paper acceptance.

Questions for authors
  1. Is the method able to reassign an object to the same instance label ID if it was occluded for a series of input frames? It would generally nice to elaborate how the temporal spawning and termination of motion trajectories works in the proposed method.

  2. What is the practical limit on the number k of track-able objects / slots? Does the number of available slots influence the results when this number is larger than the number of moving objects in the scene ? (if there are less slots than objects it obviously alters the results).

  3. How are shadows and specular reflections modeled in the reconstruction model which are frequent in urban driving scenes? Are they treated outliers or is the approach able to handle and reconstruct such effects?

6 comments
AuthorsOpenReview ↗

Response to Reviewer ZGZP (Part 1/6): Weakness 1

Weakness 1. For a method that reconstructs dynamic scenes, I would have liked to see a re-rendered video of the reconstruction results in the supplementary material in order to be able judge the quality of the reconstruction. Furthermore, a visualization of the estimated motion fields would be equally useful.

Thank you for your valuable suggestion. Building upon the reconstruction renderings and motion flow visualizations presented in Figures 3 and 4 of the main paper, we further provide videos of both in the supplementary material, offering a more comprehensive demonstration of the rendering fidelity and motion modeling capabilities of our method.

AuthorsOpenReview ↗

Response to Reviewer ZGZP (Part 2/6): Weakness 2

Weakness 2. The discussion of method limitations would be helpful.

Thank you for your valuable suggestion. Severe occlusion and complex illumination effects remain challenging problems in dynamic scene reconstruction, particularly for feed-forward methods. While our TrajForward introduces designs that help mitigate these challenges, there is still room for further improvement. We have discussed some limitations in the appendix and further elaborate on them below, including the issue of shadows and specular reflections raised in Question 3.

(1) Robustness to severe occlusion. Robust motion estimation under severe occlusion remains a fundamental challenge in dynamic scene reconstruction due to the lack of reliable visual observations. TrajForward alleviates this challenge through entity-conditioned trajectories, which encourage temporally consistent motion modeling and help maintain coherent motion estimates under partial occlusion. Nevertheless, prolonged or severe occlusion can still introduce ambiguity in entity association, potentially leading to inaccurate slot assignments and reduced reconstruction fidelity for fine-grained dynamic details. In future work, we plan to explore occlusion-aware entity association and longer-range temporal reasoning to further improve robustness under such challenging conditions.

(2) Modeling of shadows and specular reflections (also addressing Question 3). Accurately reconstructing shadows and specular reflections remains a common challenge for feed-forward reconstruction methods, including STORM, DGGT, and ReconDrive, as these effects involve complex interactions among geometry, materials, and illumination. Although TrajForward does not explicitly disentangle these factors, its Gaussian-based photometric reconstruction can implicitly capture certain shadow and reflection effects from observed appearances. However, reconstruction fidelity may still degrade under complex or changing illumination. To further address this limitation, we plan to investigate explicit illumination and reflectance decomposition, together with shadow-aware rendering, drawing inspiration from UrbanIR [1] and Relightable 3D Gaussians [2].

[1] Lin, Chih-Hao, et al. "Urbanir: Large-scale urban scene inverse rendering from a single video." 2025 International Conference on 3D Vision (3DV). IEEE, 2025.
[2] Gao, Jian, et al. "Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing." European conference on computer vision (ECCV). Cham: Springer Nature Switzerland, 2024.

AuthorsOpenReview ↗

Response to Reviewer ZGZP (Part 3/6): Weakness 3

Weakness 3. The term "4D Gaussian splatting" is a misnomer since temporally varying 3D Gaussians are not 4D Gaussians (Gaussians are well-defined in any dimension with an n-dimensional mean vector and n x n dimensional covariance matrix). The community should do an effort to avoid to use of incorrect/confusing terminology and not further propagate such naming mistakes. There are better alternatives, e.g. "dynamic Gaussian splatting".

Thank you for pointing this out. We will revise the terminology "4D Gaussian splatting" to "dynamic Gaussian splatting" in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.