VCLab

 
RESEARCH AREAS   PEOPLE   PUBLICATIONS   COURSES   ABOUT US
Home / Publications

indicator

International Conference on 3D Vision 2026
 
Splat-based Gradient-domain Fusion for Seamless View Transition
 
 
  Dongyoung Choi   Jaemin Cho   Woohyun Kang  
  Hyunho Ha   James Tompkin   Min H. Kim  
       
  KAIST Brown University  
         
  teaser
  Alleviating overfitting in sparse settings for 3D Gaussian splatting. Given sparse input, Gaussians are adjusted to match specific input training views, resulting in unreliable geometry and inconsistent color appearance through novel view transitions. We integrates gradient-domain fusion (GF) of reprojected input views at intermediate virtual poses into the GS rendering framework. This creates reference gradient fields to regularize the rendered result, producing a smoother appearance variation during view transitions. Our method achieves higher-fidelity and perceptually pleasing view synthesis in sparse scenarios than existing work.  
     
     
   
  Abstract
   
  In sparse novel view synthesis with few input views and wide baselines, existing methods often fail due to weak geometric correspondences and view-dependent color inconsistencies. Splatting-based approaches can produce plausible results near training views, but they frequently overfit and struggle to maintain smooth, realistic appearance transitions in novel viewpoints. We introduce a splat-based gradient-domain fusion method that addresses these limitations. Our approach first establishes reliable dense geometry via two-view stereo for stable initialization. We then generate intermediate virtual views by reprojecting input images, which provide reference gradient fields for gradient-domain fusion. By blending these gradients, our method transfers low-frequency, view-dependent colors to the rendered Gaussians, producing seamless appearance transitions across views. Extensive experiments show that our approach consistently outperforms state-of-the-art sparse Gaussian splatting methods, delivering robust and perceptually plausible view synthesis. A comprehensive user study further confirms that our results are perceptually preferred, with significantly smoother and more realistic color transitions than existing methods.
   
  BibTeX
  
@InProceedings{Choi:3DV:2026,
  author  = {Dongyoung Choi and Jaemin Cho and 
             Woohyun Kang and Hyunho Ha and James Tompkin
             and Min H. Kim},
  title   = {Splat-based Gradient-domain Fusion 
             for Seamless View Transition},
  booktitle = {Proc. Int. Conf. 3D Vision (2026)},
  address = {Vancouver, BC, Canada},
  year = {2026},
  month = {03},
}

   
   
icon
 Preprint paper:
PDF (2.8MB)
 

Hosted by Visual Computing Laboratory, School of Computing, KAIST.

 KAIST