acceptodds
Withdrawn from ICLR 2027

CamoSplat: Temporal Appearance Modeling for 3D Gaussian Splatting in Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Driving scenes contain local appearance changes over time, from gradual shading variations to abrupt traffic-light transitions. Capturing these changes with 3D Gaussian Splatting (3DGS) calls for temporal appearance control at the primitive level, yet current methods either keep per-Gaussian appearance fixed or tie its change to discrete per-frame codes. We introduce CamoSplat, which models a temporal appearance pattern for each Gaussian and controls how strongly it is expressed over time. These patterns come from a Mixture-of-Temporal-Bases (MoTB) that combines shared temporal functions with per-Gaussian weights, avoiding the overhead of separate per-Gaussian networks. To control when and how strongly each pattern acts, Morlet wavelet temporal gating modulates its amplitude across multiple temporal scales. The same wavelet control extends to visibility, where sensor-specific opacity wavelets add a temporal offset to a shared base opacity. CamoSplat achieves state-of-the-art photometric and LiDAR reconstruction on our temporal-variation splits from Waymo, Argoverse 2, and nuScenes, and the best overall results on standard benchmark scenes. Beyond these metrics, it precisely captures temporally localized appearance changes that prior methods collapse into ghosting or blur. We will release our code publicly.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D Gaussian Splatting, Novel View Synthesis, Autonomous Driving, Temporal Appearance Modeling, Sensor Simulation

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_aqcLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CamoSplat, which modeling temporal appearance changes in driving-scene 3D Gaussian Splatting. It introduces three main components: MoTB models different appearance changes for each Gaussian. Morlet wavelet gating controls when and how strongly these changes happen. Sensor-specific opacity wavelets model different visibility changes for camera and LiDAR. Experiments on multi datasets show improvements in image and LiDAR reconstruction quality over existing methods.

Critical strengths

S1. MoTB combines shared temporal networks with per-Gaussian weights, allowing different temporal appearance patterns without separate networks for each Gaussian.

S2. Morlet wavelet gating and sensor-specific opacity provide flexible temporal control for camera and LiDAR rendering. The method achieves good results across multiple driving datasets.

Critical weaknesses

W1. The experiments show better overall reconstruction, but do not fully verify the claimed temporal accuracy.

  1. Better image quality does not mean accurate traffic-light switching. Fig. 4 illustrates one transition, but light-state accuracy and switching-time error across events are not reported. Thus, the claim of accurate event timing remains insufficiently validated.
  2. Fig. 3 compares Morlet and Fourier using dictionaries of the same size, but shows only two visual examples. Without quantitative results, it is unclear whether Morlet is consistently better.
  3. The 50% uniform split may still place test frames close to training frames. This alone does not support the claim that simple interpolation cannot recover temporal changes.

W2. The need for the two-part temporal design and its added cost is not sufficiently established.

  1. Table 3 shows that adding gating to MoTB improves results, but it also adds parameters. It is unclear whether the gain comes from the two-module design or from more parameters. A comparison with a single temporal module of similar size would help clarify this.
  2. Training time and peak memory are about 1.8× and 2.4× those of SplatAD, while rendering speed drops from 37.4 to 21.0 FPS. The cost of each module is not reported, and there is no comparison with simpler models under similar budgets. It is unclear whether a simpler model could achieve similar quality at lower cost.
Questions for authors

Q1. Table 2 reports Depth and CD in m, while SplatAD uses median squared depth error, which has units of m². Does this paper use a different metric, or is this a labeling error?

Q2. The authors manually select 10 scenes from each dataset. How can the authors ensure that this selection does not introduce bias?

Q3. For the 50% LINSPACE split, what is the actual time gap between training and test frames? How does this setting demonstrate that simple temporal interpolation is insufficient?

Reviewer_EQ7mOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CamoSplat, an extension of SplatAD for modeling time-varying appearance and visibility in driving scenes. MoTB combines shared learned temporal networks using per-Gaussian weights. A separate gate, constructed from a fixed Morlet wavelet dictionary, modulates the resulting appearance features. The method also introduces separate temporal opacity coefficients for camera and LiDAR rendering. Experiments on temporal-variation splits from Waymo, Argoverse 2, and nuScenes report improvements in reconstruction quality, with additional results on the standard SplatAD scene splits.

Critical strengths
  1. The paper addresses an important problem in driving scene reconstruction. Modeling localized appearance changes at the Gaussian level is a reasonable approach, particularly for traffic lights and other short-term changes.
  2. The experimental results are also promising. The improvements over SplatAD, especially on nuScenes, are substantial, and the ablation study provides useful evidence for the contribution of each component.
Critical weaknesses
  1. The paper emphasizes recovering localized appearance changes at the correct timestamps, but Tables 1–2 mainly measure overall reconstruction quality rather than event-state or transition-time accuracy. Figure 4 examines one traffic-light transition, and Figure 5 and the supplementary figures provide further qualitative evidence, but none establishes how reliably the correct states are recovered across multiple events. Moreover, the crossing of internal gates in Figure 4 is not itself a measure of the rendered transition time, since Eq. (6) combines the gate with MoTB modulation. The evaluation split also raises a related concern. While the paper emphasizes wide temporal gaps, a 50% LINSPACE split does not necessarily place evaluation frames far from training observations; held-out frames may remain immediately adjacent to training frames. The actual nearest-training-frame time gaps are not reported. Consequently, although the overall reconstruction improvements are convincing, the experiments do not sufficiently establish the temporal accuracy and generalization capability that motivate the proposed method.

  2. Continuous-time appearance modeling has already been explored using Fourier representations, including StreetGS. The paper argues that Morlet gating provides an important advantage through temporal locality. However, although Sec. 3.3 describes a same-sized Fourier alternative, Figure 3 provides only selected qualitative examples, without aggregate quantitative results. Table 3 demonstrates the benefit of adding temporal gating, and Table S7 in the supplementary material compares different wavelet families, but neither isolates the advantage of Morlet over the Fourier alternative central to the paper's motivation. Furthermore, the basis-selection explanation in supplementary Sec. D contains technical inaccuracies. Since the dictionary is fixed, Haar's non-differentiability with respect to time does not imply undefined derivatives with respect to the learned coefficients (Eq. 5). Similarly, Eq. (2) defines same-scale Morlet bases as translated copies, contrary to the supplementary explanation. These issues do not invalidate the reported reconstruction improvements, but leave the necessity and specific advantage of the proposed temporal representation insufficiently justified. This limits my assessment of its technical contribution beyond existing continuous-time appearance models. Furthermore, Dynamic 3DGS has been widely studied beyond autonomous driving. Could the authors clarify which aspects of CamoSplat are generally applicable to dynamic 3DGS and which depend on driving-specific assumptions, such as tracked object trajectories and joint camera–LiDAR supervision? A clearer discussion of its relationship to general dynamic 3DGS methods would help establish the scope and significance of the technical contribution.

Questions for authors
  1. Which camera calibrations, ego poses, and object trajectories are assumed known, estimated, or optimized? Please clarify their provenance and what information is available at held-out timestamps.
  2. Is LiDAR from camera-held-out timestamps used for training, and is the same hold-out policy applied to the compared methods? Please clarify how the camera and LiDAR training/evaluation splits are coordinated.
  3. What supports the claim of "inherent self-regularization" in Sec E in Supp beyond the single-scene histogram in Fig. S2? Temporal locality and approximately zero-mean bases alone do not impose an amplitude bound when their coefficients are unconstrained.
Reviewer_iyrVOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed a framework for handling time-varying appearance in Gaussian representations of autonomous driving scenes. Results on selected sequences from Waymo, Argoverse 2, and nuScenes indicate improvements in image reconstruction and LiDAR rendering.

Critical strengths

The reported metrics show favorable results on all three datasets, with especially clear gains on nuScenes. The visualizations in Figures 4 and 5 further suggest that the method can reproduce appearance changes occurring within specific time intervals.

Critical weaknesses
  1. The evaluation covers just 30 manually chosen sequences and adopts a 50% temporal partition. As timestamps and camera perspectives vary together, the observed improvements cannot be clearly attributed to temporal appearance modeling alone. Furthermore, no numerical analysis is provided to assess how accurately the method captures event transitions.

  2. The specific innovations introduced by CamoSplat's temporal modulation mechanism remain difficult to distinguish from earlier work.

Questions for authors

Refer to the Critical Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.