Abstract
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. It achieves the highest metrics in 38 of 42 per-scene comparisons while removing windshield reflections and preserving the transmitted scene.
Method
Given only posed driving frames with windshield reflections, RIDE-Splat separates the superimposed signals in the coordinate systems where they remain consistent. A 3D Gaussian Splatting (3DGS) scene represents the parallax-bearing transmission, while one shared, signed RGB image represents the dominant camera-attached reflection. The signed image models both added reflected radiance and attenuation of the transmitted scene without clean training images, reflection masks, pretrained networks, view conditioning, or per-frame coefficients.
In Phase I, a geometry warm-up establishes the scene before joint photometric optimization recovers the 3DGS model and reflection image; an L1 penalty on the reflection biases ambiguous content toward the 3D scene. Phase II takes the maximum absolute response across reflection channels, smooths and normalizes it into a continuous spatial weight, then restarts the decomposition from scratch with that support frozen. This self-localized refit suppresses weak scene leakage without adding optimized parameters. At inference, RIDE-Splat discards the reflection image and renders the clean scene through the standard 3DGS path.
The RIDE-Splat Dataset is the first 3D reconstruction benchmark for reflection removal from driving video with physical windshield reflections and non-reflective evaluation images. Each scene contains two jointly registered traversals captured minutes apart with the same rigid camera mount: models train only on the reflective traversal (left), while the non-reflective evaluation traversal (right) is held out for evaluation. Because traffic, illumination, and the exact trajectory can differ between passes, the evaluation images are close references rather than pixel-exact targets.
Results
BibTeX
@misc{Hayes2026RIDESplat,
author = {Hayes, Kevin David and Hanson, Alex and Tu, Allen and Asthana, Pranav and Zwicker, Matthias and Goldstein, Tom},
title = {{RIDE-Splat}: Reflection Removal for In-the-Wild Driving Environments},
year = {2026},
eprint = {2610.34632},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.34632}
}Related Work
For additional papers on 3D reconstruction from degraded imagery, see our group’s work below.