Scholarlink

<p>3D Gaussian Splatting (3DGS) is an efficient scene representation and rendering method for multi-view images. However, in real-world road scenes, movable occluders often obscure the road surface and introduce occlusion-related artifacts into the reconstructed scene. To address this issue, we propose RoadGaussian, a 3DGS-based framework for batch removal of movable occluders in road scenes. The method decomposes a complete textual prompt into multiple simpler prompts and uses Grounded-Segment-Anything (G-SAM) to generate text-conditioned masks for different occluder categories. These masks are then combined with Large Mask Inpainting (LaMa) to remove occluders and recover the corresponding road regions. A loop-processing strategy further enables the automatic handling of multiple occluder categories. The processed images are incorporated into 3DGS as supervision for Gaussian optimization. This reduces the influence of road occluders on road-scene reconstruction. Experimental results on real outdoor UAV road sequences show that the proposed method can reduce visible occlusion-related artifacts under the evaluated settings. The real outdoor performance is mainly supported by qualitative comparisons and no-reference evaluation. Compared with vanilla 3DGS, the proposed method improves the mean PS-score from 0.95 to 0.97 and reduces the mean BRISQUE from 40.37 to 37.04 on the outdoor dataset. For reference-based quantitative evaluation on the controlled indoor tabletop proxy dataset, compared with GaussianEditor, the proposed method improves the average PSNR by 0.66 dB and increases VMAF and FVVDP by 11.28% and 28.64%, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RoadGaussian: A 3D gaussian splatting-based framework for road reconstruction and batch removal of occlusions

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Zhengkun Xiong,
  • Xiaolong Li,
  • Zhongxiang Yang,
  • Yirong Huang,
  • Jianhui Li

摘要

3D Gaussian Splatting (3DGS) is an efficient scene representation and rendering method for multi-view images. However, in real-world road scenes, movable occluders often obscure the road surface and introduce occlusion-related artifacts into the reconstructed scene. To address this issue, we propose RoadGaussian, a 3DGS-based framework for batch removal of movable occluders in road scenes. The method decomposes a complete textual prompt into multiple simpler prompts and uses Grounded-Segment-Anything (G-SAM) to generate text-conditioned masks for different occluder categories. These masks are then combined with Large Mask Inpainting (LaMa) to remove occluders and recover the corresponding road regions. A loop-processing strategy further enables the automatic handling of multiple occluder categories. The processed images are incorporated into 3DGS as supervision for Gaussian optimization. This reduces the influence of road occluders on road-scene reconstruction. Experimental results on real outdoor UAV road sequences show that the proposed method can reduce visible occlusion-related artifacts under the evaluated settings. The real outdoor performance is mainly supported by qualitative comparisons and no-reference evaluation. Compared with vanilla 3DGS, the proposed method improves the mean PS-score from 0.95 to 0.97 and reduces the mean BRISQUE from 40.37 to 37.04 on the outdoor dataset. For reference-based quantitative evaluation on the controlled indoor tabletop proxy dataset, compared with GaussianEditor, the proposed method improves the average PSNR by 0.66 dB and increases VMAF and FVVDP by 11.28% and 28.64%, respectively.