MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes

Yifei Wang; Wei Tian; Wenhua Bao; Jia Zhou; Hanshi Wang; Fudong Ge; Jianhua Wu; Shuai Yuan

cs.CVOct 18, 2026

MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes

Authors: Yifei Wang, Wei Tian, Wenhua Bao, Jia Zhou, Hanshi Wang, Fudong Ge, Jianhua Wu, Shuai Yuan,

Organizations: Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University · Zaofu Intelligent Technology Co., Ltd. · College of Automotive and Energy Engineering, Tongji University · Institute of Automation, Chinese Academy of Sciences ·

Abstract

Driving-scene reconstruction via feed-forward Gaussian splatting from sparse multi-view images remains challenging due to two factors. First, cross-camera photometric inconsistency, caused by variations in exposure, ISP, and illumination, can be incorrectly absorbed into Gaussian geometry. Second, unreliable local geometric supervision arises from normal-based constraints near depth discontinuities. To address these issues, we present MCSplat, a multi-view consistent feed-forward Gaussian splatting framework for driving scenes. MCSplat constructs a canonical appearance space from depth-guided multi-camera correspondences to decouple geometry learning from camera-specific appearance. A multi-scale Bilateral Grid Transformer restores raw camera appearance from canonical renderings, enabling photometric supervision without contaminating Gaussian geometry. To improve local geometric reliability, we estimate Fisher-information-based uncertainty from photometric sensitivity and use it to adaptively weight surface-oriented self-supervision. Extensive experiments demonstrate improved reconstruction quality and generalization across driving datasets. Geometry-oriented evaluations further reveal enhanced cross-view depth consistency and local surface-normal agreement, while qualitative results show fewer geometric artifacts and more coherent structures across overlapping views.

Figures & tables

Figure 2: Overview of MCSplat. (a) Depth-guided canonical appearance construction aligns inconsistent multi-view images. (b) A multi-scale bilateral-grid transformer restores camera-specific appearance from canonical renderings. (c) Fisher information-based uncertainty guides geometric self-supervision in under-constrained regions.

Figure 3: Visualization of photometric-loss uncertainty maps for Gaussian color, scale, and center parameters. The center-based map captures local structure around the white truck boundary and the jeep grille more distinctly than the scale map and the color map.

Method	Views	Waymo						nuScenes
		Reconstruction			Novel View Synthesis			Recon&NVS
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
EvolSplat ∗ ( Miao et al., 2025 )	Front	23.35	0.70	0.29	–	–	–	–	–	–
UniSplat ∗ ( Shi et al., 2026 )	Front	28.93	0.86	0.18	27.34	0.80	0.20	–	–	–
Ours	Front	31.05	0.89	0.12	27.64	0.84	0.16	–	–	–
PixelSplat ∗ ( Charatan et al., 2024 )	Multi	–	–	–	–	–	–	21.51	0.616	0.372

Table 1: Quantitative comparison of image quality on the Waymo and nuScenes datasets. The best , second-best , and third-best results are highlighted. ∗ : Reported by UniSplat. 3 : Evaluated on the front 3 views only.

Table 4

Figure 4: Qualitative comparisons on the Waymo dataset. Compared with existing methods, MCSplat produces finer and cross-camera-consistent results, avoiding texture-induced depth ambiguity within the yellow road sign.

Figure 5: Qualitative results under cross-camera appearance inconsistency. Zoom in for best view. Compared with UniSplat (c), our canonical rendering removes camera-boundary artifacts (b), while achieving effective appearance recovery (d).

Method	Eval.	Depth Reproj. (m)		Reconstruction			Pseudo-Normal Consistency
Method	Eval.	Mean $\downarrow$	RMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Mean ( ∘ ) $\downarrow$	$<5^{\circ}$ (%) $\uparrow$	$<11.25^{\circ}$ (%) $\uparrow$	$<30^{\circ}$ (%) $\uparrow$
Baseline	Raw	2.378	6.427	29.58	0.86	0.17	115.05	0.02	0.10	1.18
Baseline + CAC	Can.	0.967	4.051	31.06	0.88	0.13	97.25	0.17	0.74	4.08
Baseline + CAC + BGT	Raw	0.967	4.051	30.76	0.88	0.13	97.25	0.17	0.74	4.08
Baseline + CAC + Geo	Can.	0.910	3.920	31.09	0.88	0.14	92.82	4.24	10.08	17.38
Baseline + CAC + Geo	Raw	0.910	3.920	29.35	0.86	0.14	92.82	4.24	10.08	17.38

Table 4: Component ablation on the Waymo dataset. “Can.” and “Raw” denote canonicalized and raw GT images used for evaluation, respectively. Depth reprojection error is evaluated in adjacent-camera overlap regions, while normal consistency is measured against MoGe-2 ( Wang et al., 2025e ) pseudo-normals.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Intuitive pipeline comparison between our method and existing GFM-based feed-forward driving Gaussian frameworks.

scene_id	seg_name	scene_id	seg_name	scene_id	seg_name
008	seg_107139	054	seg_143831	120	seg_283472
009	seg_108375	059	seg_146633	122	seg_303925
012	seg_110487	060	seg_146873	129	seg_366934
013	seg_110533	065	seg_150215	133	seg_419577
014	seg_113566	067	seg_150569	134	seg_424653
016	seg_114061	072	seg_154962	137	seg_442641

Appendix

Table 5: Waymo Multi-View Appearance-Inconsistent Subset.

Method	Views	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
UniSplat	Front	29.18	0.89	0.15
Ours	Front	32.07	0.90	0.12
UniSplat	Multi	29.28	0.86	0.18
Ours	Multi	30.88	0.88	0.14

Appendix

Table 6: Quantitative reconstruction results on the Waymo Appearance-Inconsistent Subset.

Figure 7: Additional qualitative comparisons under severe cross-camera exposure inconsistency. Our method reduces exposure-induced cross-camera artifacts and alleviates texture-induced depth errors. Around the pickup truck, it avoids confusing the window depth with the background depth, yielding a more plausible depth estimate.

Figure 8: Additional qualitative comparisons. Our method effectively removes glass-surface artifacts in camera-boundary regions. Compared with LiDAR depth, it also yields improved depth estimation in pedestrian regions. On the planar poster surface, our method avoids texture-induced depth ambiguity, producing a more reasonable depth prediction.

Figure 9: Additional qualitative comparisons on the Waymo dataset. Our method reduces cross-camera stitching artifacts and produces more reasonable depth estimates, especially around road boundaries and tree or shrub boundaries.

Figure 10: Qualitative comparison of single-frame Gaussians under a novel view. Compared with the baseline, our method reconstructs cleaner road boundaries and sharper scene geometry. The highlighted regions show that the baseline suffers from geometry–appearance ambiguity caused by multi-view photometric inconsistency, leading to broken road edges, blurred boundary structures, and incomplete ground reconstruction.

Figure 11: Qualitative comparison of single-frame Gaussians under a novel view in urban scenes. The highlighted regions show that our method recovers dynamic pedestrians more faithfully than the baseline, while preserving clearer wall textures and more complete ground reconstruction.

Figure 12: Qualitative comparison of single-frame Gaussians under a novel view. UniSplat tends to confuse texture-rich regions with geometric variations, leading to blurred and distorted wall surfaces. In contrast, our method preserves sharper texture details and more consistent wall geometry in the highlighted regions.

Figure 13: Qualitative cross-lane novel-view synthesis under a 1.8 m leftward camera-trajectory shift. (a) Original trajectory; (b) shifted view rendered by UniSplat; (c) shifted view rendered by MCSplat. The highlighted regions show improved preservation of the left-foreground car and the parked motorcycles along the right sidewalk.

Figure 14: Qualitative cross-lane novel-view synthesis under a 2.0 m leftward camera-trajectory shift. (a) Original trajectory; (b) shifted view rendered by UniSplat; (c) shifted view rendered by MCSplat. Compared with UniSplat, MCSplat better preserves the highlighted pickup-truck rear structures and the red foreground car, although off-trajectory artifacts remain.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 955 papers

ConFixGS: Learning to Fix Feedforward 3D Gaussian Splatting with Confidence-Aware Diffusion Priors in Driving Scenes
May 10, 2026Rui Song, Tianhui Cai, Markus Gross +53D Gaussian Splatting Diffusion Guidance
University of California, Los Angeles · Technical University of Munich · University of Cambridge
SpectralSplat: Appearance-Disentangled Feed-Forward Gaussian Splatting for Driving Scenes
Apr 3, 2026Quentin Herau, Tianshuo Xu, Depu Meng +53D Scene Relighting 3D Gaussian Splatting
Applied Intuition · The Hong Kong University of Science and Technology · University of California, Berkeley
PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations
May 12, 2026Cheng Chi, Xianqi Wang, Hongcheng Luo +93D Reconstruction Multi-View 3D Reconstruction
Xiaomi EV · Huazhong University of Science and Technology · Zhejiang University

May 10, 2026J/K move · Enter open · S save

ConFixGS: Learning to Fix Feedforward 3D Gaussian Splatting with Confidence-Aware Diffusion Priors in Driving Scenes

Rui Song, Tianhui Cai, Markus Gross, Xingcheng Zhou+4

University of California, Los Angeles · Technical University of Munich · University of Cambridge

3D Gaussian Splatting Diffusion Guidance Sparse-View 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

Feedforward 3D Gaussian Splatting (3DGS) often struggles in trajectory-based sparse-view driving scenes. Existing Gaussian repair methods mainly target optimization-based 3DGS, while diffusion-based repair is typically restricted to iterative refinement near observed viewpoints, leaving feedforward 3DGS repair underexplored. We propose ConFixGS, a plug-and-play method that learns to fix feedforward 3DGS with confidence-aware diffusion priors. Starting from a pretrained feedforward model, ConFixGS generates diffusion-enhanced local pseudo-targets and validates them through reprojection-based cross-checking against support views. The resulting dense confidence maps guide refinement, enhancing reliable details while suppressing hallucinated or inconsistent evidence. On Waymo, nuScenes, and KITTI, ConFixGS improves challenging novel view synthesis, with PSNR gains of up to 3.68 dB and FID reduced by nearly half. Our results highlight confidence-aware fusion of generative priors and support-view consistency as a key principle for robust feedforward 3D driving scene reconstruction.

Abstract

3D Reconstruction

Explore similar work

ConFixGS: Learning to Fix Feedforward 3D Gaussian Splatting with Confidence-Aware Diffusion Priors in Driving Scenes

SpectralSplat: Appearance-Disentangled Feed-Forward Gaussian Splatting for Driving Scenes

PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations