🤖 AI Summary
This paper addresses the challenge of accurate 3D modeling of vehicle damage from single-view images. We propose a training-free 2D→3D damage segmentation method. First, camera poses are estimated via Structure-from-Motion. Then, a 3D Gaussian point lattice representing the vehicle geometry is projected onto the image plane; leveraging Z-buffering and a depth-opacity normal distribution model, we perform point cloud filtering and refine the damage mask to achieve geometrically consistent mapping from 2D damage regions to the 3D Gaussian point space. Unlike conventional approaches requiring multi-view consistency, our method precisely localizes fine-grained damage—such as scratches and dents—from a single view, significantly improving geometric accuracy and robustness in 3D reconstruction. Extensive experiments validate its effectiveness, and the implementation is publicly available.
📝 Abstract
Automatic car damage detection has been a topic of significant interest for the auto insurance industry as it promises faster, accurate, and cost-effective damage assessments. However, few works have gone beyond 2D image analysis to leverage 3D reconstruction methods, which have the potential to provide a more comprehensive and geometrically accurate representation of the damage. Moreover, recent methods employing 3D representations for novel view synthesis, particularly 3D Gaussian Splatting (3D-GS), have demonstrated the ability to generate accurate and coherent 3D reconstructions from a limited number of views. In this work we introduce an automatic car damage detection pipeline that performs 3D damage segmentation by up-lifting 2D masks. Additionally, we propose a simple yet effective learning-free approach for single-view 3D-GS segmentation. Specifically, Gaussians are projected onto the image plane using camera parameters obtained via Structure from Motion (SfM). They are then filtered through an algorithm that utilizes Z-buffering along with a normal distribution model of depth and opacities. Through experiments we found that this method is particularly effective for challenging scenarios like car damage detection, where target objects (e.g., scratches, small dents) may only be clearly visible in a single view, making multi-view consistency approaches impractical or impossible. The code is publicly available at: https://github.com/DragosChileban/CrashSplat.