RigidDriveGS: Object-Centric Vehicle Reconstruction from Fixed Multi-Camera Drive-Through Captures

Nitin Kulkarni; Akhil Devarashetti; Charlie Cluss; Livio Forte; Philip Schneider; Chunming Qiao; Alina Vereshchaka

cs.CVMar 27, 2026

RigidDriveGS: Object-Centric Vehicle Reconstruction from Fixed Multi-Camera Drive-Through Captures

Authors: Nitin Kulkarni, Akhil Devarashetti, Charlie Cluss, Livio Forte, Philip Schneider, Chunming Qiao, Alina Vereshchaka

Organizations: ACV Auctions, Buffalo, NY, USA · University at Buffalo, Buffalo, NY, USA

Abstract

High-fidelity 3D reconstruction of vehicle exteriors improves buyer confidence in online automotive marketplaces, but fixed multi-camera drive-through capture departs from standard Structure-from-Motion (SfM): static cameras observe a moving vehicle amid clutter, rotating wheels, fisheye distortion, and specular reflections. We present RigidDriveGS, an object-centric pipeline combining motion-gated tracking, bifurcated masks that exclude wheels for SfM while optionally retaining them for rendering, mask-guided dense correspondences on raw distorted images, rig-constrained global SfM, and clean object-centric neural rendering with trimap-aware opacity or density supervision and native-alpha white compositing. We evaluate 25 vehicles from 10 dealership installations and six phone-recorded vehicles, each captured from 14 independent handheld positions. Under a staggered per-camera temporal-block holdout, 3DGUT+MCMC achieves 30.42 dB PSNR, 0.90 SSIM, and 0.20 LPIPS, improving PSNR by 7.78 dB over undistorted 3D-GS; independently recorded reference LiDAR scans yield full/body Chamfer-L1 of 0.026/0.013 m and F1@5 cm of 0.703/0.885. These results support clutter-free, object-centric novel-view synthesis without controlled studio infrastructure.

Figures & tables

Figure 1: RigidDriveGS model pipeline. Raw frames are tracked and masked, dense matches are admitted on rigid vehicle endpoints, object-centric SfM estimates a vehicle-centric scaffold, and distortion-aware splatting reconstructs appearance. The synchronized fixed-rig branch uses CAD pose priors; the handheld branch replaces this block with rig-free global SfM over 14 independent passes.

	Fixed rig	Handheld transfer
Acquisition	14 fixed, synchronized cameras	14 independent handheld passes
Intrinsics	ChArUco; 8-param. fisheye	ChArUco; $k_{1}{+}k_{2}$ , zero tangential
Extrinsics	CAD priors + bounded refinement	unknown; object-centric SfM
Shared stages	masks, matching, object-centric SfM, rendering
Reconstruction scale	CAD relative translations	global similarity gauge
Surface reference	none	LiDAR, evaluation only

Table 1: Calibration and capture geometry for synchronized fixed-rig and independent handheld acquisition.

Figure 2: Fixed-rig data acquisition. Fourteen synchronized wide-angle cameras on two pillars observe a vehicle traversing a cluttered capture volume. The three vertical tiers and front-, side-, and rear-facing orientations provide complementary exterior coverage in one pass.

Figure 3: Rigid vehicle isolation and tracking. Top: representative raw frames. Bottom: geometry-valid masks produced by motion-gated tracking. The masks retain the rigid vehicle body while suppressing clutter and rotating wheels for matching and SfM.

Figure 4: Mask-guided RoMA v2 correspondences on raw distorted images. Dense matches retain full-image context while endpoint admission restricts SfM evidence to the rigid vehicle body across wide viewpoint changes.

	Fixed rig				Handheld
Method	$R\uparrow$	$E\downarrow$	$\bar{L}\uparrow$	$N$ (M) $\uparrow$	$R\uparrow$	$E\downarrow$	$\bar{L}\uparrow$	$N$ (M) $\uparrow$
SIFT	0.00	–	–	–	0.00	–	–	–
DISK+LG	0.00	–	–	–	0.00	–	–	–
LoFTR	0.00	–	–	–	0.63	1.11	3.77	0.38
RoMA v2	1.00	0.61	4.01	0.96	1.00	1.04	3.41	1.15

Table 2: Failure-aware full 14-view SfM comparison over 25 fixed-rig and six handheld vehicles.

R

: image-weighted valid registration coverage;

E

: reprojection error (px);

\bar{L}

: mean track length;

N

: mean sparse points in millions.

E

\bar{L}

, and

N

use valid reconstructions only; – denotes that no reconstruction satisfied the validity criteria.

Figure 5: Representative successful object-centric sparse reconstruction from multiple viewpoints. The coherent vehicle shell and camera poses provide the common geometric scaffold for downstream rendering.

Figure 6: Qualitative comparison of novel-view vehicle rendering across diverse scenes. The leftmost column shows ground truth images; remaining columns show vehicle-centric renders from the compared methods.

	Fixed rig			Handheld
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
3DGS	22.64	0.88	0.25	22.69	0.89	0.21
3DGUT	29.24	0.87	0.21	32.33	0.93	0.18
3DGUT+MCMC	30.42	0.90	0.20	32.71	0.94	0.17
2DGS	24.89	0.87	0.26	27.66	0.93	0.19
Fisheye-GS	20.50	0.86	0.28	22.33	0.88	0.23

Table 3: Held-out rendering on privacy-excluded rigid-body pixels using the shared staggered per-stream split. Metrics are averaged separately over the fixed-rig and handheld evaluations.

Method	Full C-L1 $\downarrow$	Body C-L1 $\downarrow$	Full F1 $\uparrow$	Body F1 $\uparrow$
3DGS	0.030	0.028	0.545	0.568
3DGUT	0.024	0.021	0.699	0.680
3DGUT+MCMC	0.026	0.013	0.703	0.885
2DGS	0.029	0.027	0.516	0.560
Fisheye-GS	0.027	0.014	0.694	0.877
Instant-NGP	0.214	0.212	0.262	0.253

Table 4: Full 14-view handheld surface comparison. Chamfer-L1 is in meters; F1 uses a 5-cm threshold.

	Fixed rig				Handheld
SfM variant	$R\uparrow$	$E\downarrow$	$\bar{L}\uparrow$	$N$ (M) $\uparrow$	$R\uparrow$	$E\downarrow$	$\bar{L}\uparrow$	$N$ (M) $\uparrow$
Full pipeline	1.00	0.61	4.01	0.96	1.00	1.04	3.41	1.15
w/o motion gate	0.55	1.03	2.69	1.60	1.00	1.04	3.22	1.12
w/o rigid endpoint masks	0.00	–	–	–	0.00	–	–	–
w/o rig priors	0.00	–	–	–	N/A
SIFT+COLMAP	0.00	–	–	–	0.00	–	–	–

Table 5: Full 14-view component ablations.

R

is image-weighted valid registration coverage.

E

\bar{L}

, and

N

use valid reconstructions only; – denotes that no reconstruction satisfied the validity criteria, and N/A denotes an inapplicable ablation. The handheld pipeline has no rig-pose prior. Rendering-mask variants are compared using held-out LPIPS.

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

RIDE-Splat: Reflection Removal for In-the-Wild Driving Environments
Oct 25, 2026Kevin David Hayes, Alex Hanson, Allen Tu +33D Reconstruction 3D Gaussian Splatting
University of Maryland, College Park
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
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

Oct 25, 2026J/K move · Enter open · S save

RIDE-Splat: Reflection Removal for In-the-Wild Driving Environments

Kevin David Hayes, Alex Hanson, Allen Tu, Pranav Asthana+2

University of Maryland, College Park

3D Reconstruction 3D Gaussian Splatting Inverse Rendering 3D Reconstruction Benchmarks

Open paper AskPDF arXiv

Driving video captured through a windshield violates the world-space consistency assumed by 3D Gaussian Splatting (3DGS): the road scene exhibits parallax while cabin reflections remain attached to the camera, causing standard 3DGS to embed reflected content into the reconstructed scene. We introduce RIDE-Splat, a pipeline for reflection-free 3D reconstruction from posed driving video without clean images, reflection masks, or pretrained networks. RIDE-Splat assigns world-consistent transmission through the windshield to a 3DGS scene and the dominant camera-attached reflection to one shared, signed RGB image. It builds the clean reconstruction in two passes. First, it learns the 3D geometry and the signed reflection image, representing both added radiance and transmission attenuation. Then, it uses the reflection-removed training images to fit a clean 3DGS model. At inference, the reflection image is discarded, preserving the standard 3DGS rendering path. We also introduce the RIDE-Splat Dataset, the first 3D reconstruction benchmark for reflection removal from driving video -- with real driving scenes, physical windshield reflections in training views, and separately captured non-reflective evaluation images. Across the RIDE-Splat Dataset, XLD, and Wayve, RIDE-Splat outperforms 3DGS, DC-Gaussian, and RobustSplat on all nine aggregate metrics, improving PSNR by up to 3.59 dB and reducing LPIPS by up to 0.0876. We achieve the highest metrics in 38 out of 42 per-scene comparisons and removes windshield reflections that remain visible in the baselines while preserving the transmitted scene.

Abstract

3D Reconstruction

Explore similar work

RIDE-Splat: Reflection Removal for In-the-Wild Driving Environments

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

PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations