- 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:
- Motion-bearing slots: Learned slots group Gaussians through hard assignments, encouraging shared motion within each group without instance annotations.
- Hierarchical continuous trajectories: Shared slot-level trajectories capture coarse motion, while Gaussian-specific residual trajectories accommodate local variations.
- 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
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."
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.
Response to Reviewer JQ23 (Part 1/3): Question 1
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
Table: Scene Flow Estimation Comparison