.png){kind=icon}
.png){kind=icon}
Return
CLIPSplat: High-quality 3D Gaussian splatting from sparse multiview images based on CLIP feature fusion
.png){kind=icon}
PRE
AI
X
G
W
W
Z
DOI：10.1117/1.JEI.35.2.023043.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The CLIPSplat proposed in this paper is a feed-forward 3D Gaussian Splatting model, which is dedicated to solving the problem of low reconstruction quality due to low-textured regions or repeated textured regions in sparse multiview image reconstruction by currently existing methods. The method innovatively fuses the global semantic prior provided by the CLIP image encoder with multiview geometric features and achieves adaptive feature complementation through the designed dynamic multichannel spatial fusion module, which significantly improves the reconstruction robustness in complex scenes. CLIPSplat can be trained end-to-end with only RGB photometric supervision and does not require scene-specific optimization. Experimental validation on RealEstate10K and ACID datasets shows that CLIPSplat outperforms existing methods in PSNR, SSIM, and LPIPS metrics, and especially demonstrates higher reconstruction quality in challenging scenes such as low-texture walls and repetitive patterns. This work provides a new paradigm for high-quality 3D reconstruction of sparse views.
Keywords:
3D Gaussian splatting
novel view synthesis
3D reconstruction
feature fusion
depth estimation
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers