Organizations: Department of Automation, Key Laboratory of System Control and Information Processing of Ministry of Education, State Key Laboratory of Avionics Integration and Aviation System-of-Systems Synthesis, Shanghai Jiao Tong University, Shanghai 200240, China.
Multi-traversal scene reconstruction is important for high-fidelity autonomous driving simulation and digital twin construction. This task involves integrating multiple sequences captured from the same geographical area at different times. In this context, a primary challenge is the significant appearance inconsistency across traversals caused by varying illumination and environmental conditions, despite the shared underlying geometry. This paper presents ADM-GS (Appearance Decomposition Gaussian Splatting for Multi-Traversal Reconstruction), a framework that applies an explicit appearance decomposition to the static background to alleviate appearance entanglement across traversals. For the static background, we decompose the appearance into traversal-invariant material, representing intrinsic material properties, and traversal-dependent illumination, capturing lighting variations. Specifically, we propose a neural light field that utilizes a frequency-separated hybrid encoding strategy. By incorporating surface normals and explicit reflection vectors, this design separately captures low-frequency diffuse illumination and high-frequency specular reflections. Quantitative evaluations on the Argoverse 2 and Waymo Open datasets demonstrate the effectiveness of ADM-GS. In multi-traversal experiments, our method achieves a +0.98 dB PSNR improvement over existing latent-based baselines while producing more consistent appearance across traversals. Code will be available at https://github.com/IRMVLab/ADM-GS.
Figures & tables
Fig. 1: Consistent Multi-Traversal Reconstruction via Appearance Decomposition. (Left) Input images from multi-traversal scenes exhibit appearance inconsistencies due to changes in illumination, weather, and time of day. (Middle) ADM-GS decomposes the static scene appearance into a traversal-invariant material field ( M ) and a traversal-dependent light field ( Lm ), enabling structured modeling of cross-traversal appearance variation. (Right) Rendering results show that ADM-GS preserves consistent material across different environmental conditions while capturing traversal-specific illumination variations, leading to improved multi-traversal novel view synthesis.
Fig. 2: Framework of Appearance Decomposition Gaussian Splatting for Multi-Traversal Reconstruction. ADM-GS represents the scene with a hybrid scene graph consisting of static, object, and sky nodes. Its core static node decomposes appearance into a traversal-invariant material field and a traversal-dependent light field, where normal and reflection-vector cues are introduced to improve illumination prediction. A traversal-conditioned opacity gating module, adapted from 4DGF [ 1 ] , uses the traversal embedding em to predict traversal-specific opacity, which suppresses geometrically inconsistent regions during training. Ground-truth RGB images and pseudo supervision on material and normals are used for optimization.
Fig. 3: Illustration of frequency-separated illumination cues. Surface normals are mainly associated with low-frequency diffuse illumination, while reflection vectors are more informative for high-frequency view-dependent specular effects.
Dataset
Method
Novel View Synthesis
Scene Reconstruction
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Argoverse 2
PVG [ 28 ]
23.04
0.782
0.290
24.37
0.824
0.275
StreetGS [ 15 ]
23.40
0.795
0.197
25.02
0.857
0.181
OmniRe [ 18 ]
23.61
0.797
0.194
25.31
0.859
0.176
Bilateral-Driving [ 21 ]
23.70
0.794
0.183
25.74
0.865
0.164
4DGF [ 1 ]
29.48
0.890
0.155
30.17
0.902
0.146
TABLE I: Quantitative comparison on single-traversal scenarios. We report results for novel view synthesis and scene reconstruction. Bold and underlined values indicate the best and second-best results, respectively.
Fig. 4: Visual comparison of novel view synthesis on dynamic scenes. We compare our method with recent approaches on the Argoverse 2 (top row) and Waymo Open (bottom row) datasets. Methods such as Bilateral-Driving [ 21 ] and OmniRe [ 18 ] show noticeable blur and ghosting artifacts in dynamic-object regions. In contrast, our method produces sharper novel-view renderings and is visually closer to the ground truth.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
4DGF [ 1 ]
26.40
0.780
0.388
MTGS [ 2 ]
23.28
0.717
0.287
ADM-GS (Ours)
27.38
0.788
0.368
TABLE II: Quantitative comparison on multi-traversal reconstruction on the Argoverse 2 dataset. The models are trained on three traversals with varying illumination conditions. Bold indicates the best performance.
Fig. 5: Qualitative analysis of appearance decomposition on the Argoverse 2 multi-traversal dataset. Each row corresponds to a different traversal of the same scene. The estimated material remains largely consistent across traversals, while the illumination maps capture traversal-dependent illumination changes. The normal maps provide geometric cues that support illumination estimation.
Fig. 6: Qualitative application: cross-traversal scene relighting. We demonstrate the controllability of the learned decomposition by transferring traversal-dependent illumination from Traversal T2 to a scene observed in Traversal T1 . Column 1 shows the source image from T1 . Column 2 shows the target traversal used to provide the illumination condition from T2 . Column 3 shows the estimated traversal-invariant material map M^T1 . Column 4 shows the relit result obtained by combining the material from T1 with the illumination condition from T2 . Observe that the shadow cast by the tree (red box) follows the illumination pattern of T2 , while the underlying road texture remains consistent with T1 . The relighting operation is applied only to the static background, while dynamic foreground objects are not edited.
Method Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
Implicit Appearance
27.10
0.782
0.383
Explicit Decomposition
27.31
0.785
0.380
+ Geometry Constraints
27.39
0.789
0.371
TABLE III: System-level ablation on Argoverse 2. We analyze the transition from implicit appearance modeling to explicit appearance decomposition on the multi-traversal NVS task. Explicit Decomposition includes pseudo material supervision, and + Geometry Constraints further adds pseudo normal supervision together with geometric cues.
Variant
Lnormal
Inputs
Metrics
r
n
PSNR ↑
SSIM ↑
LPIPS ↓
Baseline
×
×
×
28.65
0.856
0.211
+ Normal Loss
✓
×
×
28.77
0.859
0.196
+ Reflection
✓
✓
×
28.86
0.861
0.187
+ Refl. & Normal
✓
✓
✓
28.91
0.862
0.186
TABLE IV: Ablation of input features for the illumination MLP on Waymo. We analyze the impact of geometric supervision ( Lnormal ) and different input feature combinations ( r,n ) on the unseen test set.
Feed-forward 3D Gaussian Splatting methods have achieved impressive reconstruction quality for autonomous driving scenes, yet they entangle scene geometry with transient appearance properties such as lighting, weather, and time of day. This coupling prevents relighting, appearance transfer, and consistent rendering across multi-traversal data captured under varying environmental conditions. We present SpectralSplat, a method that disentangles appearance from geometry within a feed-forward Gaussian Splatting framework. Our key insight is to factor color prediction into an appearance-agnostic base stream and and appearance-conditioned adapted stream, both produced by a shared MLP conditioned on a global appearance embedding derived from DINOv2 features. To enforce disentanglement, we train with paired observations generated by a hybrid relighting pipeline that combines physics-based intrinsic decomposition with diffusion based generative refinement, and supervise with complementary consistency, reconstruction, cross-appearance, and base color losses. We further introduce an appearance-adaptable temporal history that stores appearance-agnostic features, enabling accumulated Gaussians to be re-rendered under arbitrary target appearances. Experiments demonstrate that SpectralSplat preserves the reconstruction quality of the underlying backbone while enabling controllable appearance transfer and temporally consistent relighting across driving sequences.
Quentin Herau, Tianshuo Xu, Depu Meng +5
Applied Intuition · The Hong Kong University of Science and Technology · University of California, Berkeley
High-fidelity visual reconstruction and novel-view synthesis are essential for realistic closed-loop evaluation in autonomous driving. While 4D Gaussian Splatting (4DGS) offers a promising balance of accuracy and efficiency, existing per-scene optimization methods require costly iterative refinement, rendering them unscalable for extensive urban environments. Conversely, current feed-forward approaches often suffer from degraded photometric quality. To address these limitations, we propose ReconDrive, a feed-forward framework that leverages and extends the 3D foundation model VGGT for rapid, high-fidelity 4DGS generation. Our architecture introduces two core adaptations to tailor the foundation model to dynamic driving scenes: (1) Hybrid Gaussian Prediction Heads, which decouple the regression of spatial coordinates and appearance attributes to overcome the photometric deficiencies inherent in generalized foundation features; and (2) a Static-Dynamic 4D Composition strategy that explicitly captures temporal motion via velocity modeling to represent complex dynamic environments. Benchmarked on nuScenes, ReconDrive significantly outperforms existing feed-forward baselines in reconstruction, novel-view synthesis, and 3D perception. It achieves performance competitive with per-scene optimization while being orders of magnitude faster, providing a scalable and practical solution for realistic driving simulation.
Haibao Yu, Kuntao Xiao, Jiahang Wang +7
The University of Hong Kong · Tuojing Intelligence · King’s College London +2
High-fidelity reconstruction of dynamic urban environments is a cornerstone of autonomous driving simulation and large-scale world modeling. While 3D Gaussian Splatting (3DGS) has established a new standard for real-time rendering, its reliance on expensive per-scene optimization limits scalability. Conversely, recent feedforward methods that infer Gaussian parameters offer faster speed but face fundamental bottlenecks: they are memory-prohibitive at high resolutions and struggle to fuse dense multi-view observations consistently. This paper presents L2D2-GS, a unified framework that reformulates generalizable reconstruction not as a one-shot regression, but as a robust iterative process of optimization and densification. To resolve the ambiguity of supervision in primitive generation, we propose a self-supervised densification policy that derives explicit reward signals from global reconstruction gains to guide local densification. Furthermore, we mitigate irreversible early-stage artifacts through a geometric regularization mechanism, utilizing reparameterization to constrain the optimization manifold and prevent convergence to poor local optima. Extensive experiments on the PandaSet and Waymo datasets demonstrate that our method achieves state-of-the-art reconstruction fidelity and strong zero-shot generalization, while using fewer primitives than competing baselines.
Zetian Song, Chenming Wu, Junnan Liu +6
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing, 100871, China · Xiaomi EV, Beijing, 100085, China