Multi-perspective-camera systems are widely adopted since they provide broad visual coverage for autonomous driving, motivating the synthesis of a unified wide view from spatially separated observations. In this work, we formulate wide view synthesis (WVS), which aims to synthesize a geometrically consistent wide view at a specified virtual viewpoint from multiple perspective cameras. Unlike conventional image stitching, WVS requires reasoning about 3D geometry and visibility to handle depth-dependent parallax and occlusions arising from spatially separated camera centers. While vanilla 3D Gaussian Splatting (3DGS) enables high-fidelity view synthesis, its reliance on per-scene optimization limits its applicability to driving scenes. Furthermore, WVS introduces additional challenges, e.g., peripheral stretching under perspective projection and photometric inconsistencies across cameras. To address these challenges, we propose Wide-View Gaussian Splatting (WV-GS), a feed-forward Gaussian-splatting framework that synthesizes wide views without per-scene optimization. Specifically, we introduce Piecewise Perspective Rendering (PPR) to mitigate peripheral stretching through local perspective projections sharing a common camera center, and Spatial Affine Photometric Modeling (SAPM) to account for camera-dependent photometric variations during training. We further introduce WideDrive, a synthetic benchmark providing dense ground-truth wide views and adapt a LiDAR-projected sparse-reference protocol to real-world datasets. Extensive experiments on multiple datasets demonstrate the effectiveness of WV-GS for WVS. Dataset and code will be publicly released.
Figures & tables
Figure 2: Stitching images of different methods. (a)–(c) Input images captured by three perspective cameras with limited overlap. (d)–(g) Results obtained using REwarp [ 21 ] , DSFN [ 19 ] , RopStitch [ 33 ] , and UniStitch [ 28 ] , respectively, showing severe geometric distortions and visible artifacts.
Figure 3: WF-GS overview. A feed-forward backbone predicts Gt from the input context Ct . (a) Training: SAPM maps canonical renders Ii,t to corrected renders Ii,t , which are compared with observed images Ii,t . Reference depths Di,tref supervise predicted depths Di,t through Ldep , while Lcon couples predicted and rasterized depths Di,trend . (b) Inference: PPR composes local canonical renders into the wide-FOV image Iw,tppr .
Figure 4: Architecture of SAPM. SAPM maps canonical renders Ii,t to camera-specific predictions Ii,t during training. Blue and orange blocks denote 3×3 and 1×1 convolutions, respectively. Concat: channel concatenation; Up: bilinear upsampling; ⊙ and ⊕ : element-wise multiplication and addition.
Figure 5: Illustration of PPR. Conceptual schematic of PPR illustrated with M=3 local perspective planes Πm ( m∈{1,2,3} ). All planes share the target optical center ow,t and cover adjacent angular sub-regions. Their canonical renders Iw,tm are concatenated horizontally into the wide-FOV output Iw,tppr .
Table 1: Results on nuScenes [ 3 ] with the best and second-best results highlighted.
Method
HM-PSNR ↑
HM-SSIM ↑
HM-MAE ↓
HM-RMSE ↓
Success (%) ↑
Non-learnable
AutoStitch
13.85
0.865
39.48
60.59
96.11
APAP
11.21
0.776
69.88
90.00
0.24
Learnable
UDIS2
12.42
0.819
55.44
76.43
72.78
REwarp
10.53
0.759
72.39
92.90
7.46
Table 2: Results on Lyft [ 26 ] with the best and second-best results highlighted.
Method
PSNR ↑
SSIM ↑
MAE ↓
RMSE ↓
LPIPS ↓
Success (%) ↑
Non-learnable
AutoStitch
11.04
0.494
69.97
89.79
0.463
79.62
APAP
11.96
0.539
48.96
68.87
0.441
37.62
Learnable
UDIS2
15.21
0.605
28.97
46.12
0.366
98.20
REwarp
6.67
0.420
83.98
123.03
0.558
19.28
Table 3: Results on WideDrive with the best and second-best results highlighted.
Figure 7: Visual comparison of WVS on nuScenes [ 3 ] . (a)–(c) Input images from three overlapping perspective cameras. (d)–(j) Wide-FOV outputs from RopStitch [ 33 ] , REwarp [ 21 ] , UniStitch [ 28 ] , StabStitch++ [ 32 ] , DSFN [ 19 ] , AutoStitch [ 2 ] , and WV-GS (ours), respectively.
Figure 8: Visual comparison of the SAPM ablation. (a)–(c): input images. (d)–(f) correspond to WV-GS w/o SAPM, WV-GS w/o SAPM + OpenCV and full WV-GS. Without SAPM, the synthesized wide views inherit the per-camera color and exposure offsets, which remain visible as brightness discontinuities and color seams in the overlapping regions, while the OpenCV-based preprocessing only partially suppresses them. SAPM instead models the camera-dependent photometric response during training and yields a consistent appearance across the whole wide field of view.
Settings
HM-PSNR ↑
HM-SSIM ↑
HM-MAE ↓
HM-RMSE ↓
w/o SAPM
23.32
0.979
9.66
18.47
w/o SAPM + OpenCV
23.45
0.978
9.71
18.31
w/o PPR
22.42
0.969
11.96
20.97
Full (Ours)
23.76
0.977
9.62
17.69
Table 4: Ablation on core modules. “w/o SAPM”: trained without photometric modeling; “w/o SAPM + OpenCV”: SAPM replaced by OpenCV-based color preprocessing; “w/o PPR”: a single wide-angle perspective projection replaces PPR.
M
θ
HM-PSNR ↑
HM-SSIM ↑
HM-MAE ↓
HM-RMSE ↓
3
∼80∘
23.76
0.977
9.62
17.69
5
∼45∘
23.42
0.972
10.91
19.19
7
∼35∘
23.42
0.971
11.10
19.37
9
∼25∘
23.48
0.971
11.12
19.35
Table 5: Ablation on the number of projection planes M in PPR.
L
HM-PSNR ↑
HM-SSIM ↑
HM-MAE ↓
HM-RMSE ↓
1
23.35
0.976
10.20
18.46
2
23.50
0.977
9.94
18.14
3
23.76
0.977
9.62
17.69
4
23.58
0.977
9.83
17.97
5
23.54
0.977
9.87
18.06
Table 6: Ablation on the temporal context length L .
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Dataset composition of WideDrive. WideDrive contains both daytime and night time scenes for better diversity. At each timestamp, we capture three perspective images (a)–(c)/(e)–(g) from spatially separated cameras, together with a wide-view image (d)/(h) that serves as the ground truth. Images (a)/(c)/(e)/(g) are additionally rendered with an exposure offset of −1 EV, emulating the auto-exposure mismatch of real cameras. LiDAR point clouds are also recorded but are omitted here for clarity.
Dataset
Type
Source cameras
Input FOV (overlap)
Exposure mismatch
Wide-view GT
Eval. protocol
nuScenes [ 3 ]
Real
3 persp. ( 1600×900 )
∼140∘ ( ∼20∘ )
Natural
× (Sparse LiDAR)
LPSR (Sparse)
Lyft [ 26 ]
Real
3 persp. ( 1920×1080 )
∼140∘ ( ∼30∘ )
Natural
× (Sparse LiDAR)
LPSR (Sparse)
WideDrive (Ours)
Synthetic
3 persp. ( 1920×1080 )
132∘ ( 24∘ / 40%)
Controlled ( −1 EV)
✓ ( 120∘ Dense)
Dense (Full Image)
Appendix
Table 7: Comparison of benchmark datasets for WF-NVS. Unlike existing real-world driving benchmarks that lack ground-truth wide-angle perspective captures, WideDrive provides native, dense, seam-free wide-view ground-truth images alongside calibrated perspective source views and realistic cross-camera exposure offsets.
Figure 10: Illustration of the LPSR protocol. Given the captured perspective images (a)–(c) and their calibrated LiDAR sweeps, we project the colored LiDAR point cloud onto (d) a single wide-angle perspective imaging plane and (e) our M piecewise perspective planes of PPR, yielding a sparse reference panel and its validity mask on which all metrics are computed. Only pixels hit by projected points are evaluated; black regions carry no LiDAR evidence and are excluded from the scores.
Figure 11: Visual comparison of WVS on Lyft [ 26 ] . (a)–(c) Input images from three overlapping perspective cameras. (d)–(k) Wide-view outputs from DSFN [ 19 ] , RopStitch [ 33 ] , UDIS2 [ 31 ] , REwarp [ 21 ] , StabStitch++ [ 32 ] , AutoStitch [ 2 ] , UniStitch [ 28 ] , and WF-GS (ours), respectively. APAP [ 39 ] fails on this scene.
Figure 12: Rendering results of WVS with different number of projection planes M in PPR. (a)–(c) Input images. (d)–(g) Images rendered by WF-GS with M=3/5/7/9 .
A fundamental bottleneck in Novel View Synthesis (NVS) for autonomous driving is the inherent supervision gap on novel trajectories: models are tasked with synthesizing unseen views during inference, yet lack ground truth images for these shifted poses during training. In this paper, we propose VisionNVS, a camera-only framework that fundamentally reformulates view synthesis from an ill-posed extrapolation problem into a self-supervised inpainting task. By introducing a ``Virtual-Shift'' strategy, we use monocular depth proxies to simulate occlusion patterns and map them onto the original view. This paradigm shift allows the use of raw, recorded images as pixel-perfect supervision, effectively eliminating the domain gap inherent in previous approaches. Furthermore, we address spatial consistency through a Pseudo-3D Seam Synthesis strategy, which integrates visual data from adjacent cameras during training to explicitly model real-world photometric discrepancies and calibration errors. Experiments demonstrate that VisionNVS achieves superior geometric fidelity and visual quality compared to LiDAR-dependent baselines, offering a robust solution for scalable driving simulation.
Hongbo Lu, Liang Yao, Chenghao He +4
Shanghai Jiao Tong University · COWARobot Co. Ltd. · Hohai University
While recent feed-forward 3D Gaussian Splatting methods enable efficient driving-scene reconstruction from surround-view images, their reliance on a single observation fundamentally hinders novel-view synthesis. Regions occluded in the input can become visible from novel viewpoints, leaving the reconstructed Gaussians under-constrained and incomplete. Moving beyond image-space inpainting for rendered views, we introduce OccSplat, a feed-forward framework whose core contribution is Occlusion-Aware Gaussian Densification, explicitly targeting missing scene geometry rather than missing pixels. By predicting new Gaussian primitives for unsupported regions, OccSplat completes the scene directly in the 3D representation. Specifically, OccSplat first identifies occluded regions that lack Gaussian support through Gaussian Occlusion Mapping. Gaussian Occlusion Completion then generates new primitives for these regions. The predicted primitives are integrated into the original representation to improve novel-view synthesis while retaining feed-forward inference without per-scene optimization. Experiments on the nuScenes benchmark demonstrate that OccSplat improves on the compared methods and yields particularly large gains in occluded regions, leading to more accurate and visually coherent driving-scene reconstruction.
Dung Do Tien, Kien Nguyen Do Trung, Kiet Dang Vu +2
AITech Lab., Ho Chi Minh City University of Technology (HCMUT) · Vietnam National University Ho Chi Minh City (VNUHCM)
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