OGGSplat: Open Gaussian Growing for Generalizable Reconstruction with Expanded Field-of-View

Yanbo Wang*, Ziyi Wang*, Wenzhao Zheng, Jie Zhou, Jiwen Lu†
Tsinghua University
* Equal contributions. † Corresponding author.
Paper arXiv Code

OGGSplat is a novel open-vocabulary generalizable 3D reconstruction framework
that extends the field-of-view from sparse input images by leveraging semantic priors.

Demo: OGGSplat for extrapolated open-vocabulary query

OGGSplat can extrapolate the Gaussians beyond the input view by leveraging semantic priors from diffusion models.

Overview

OGGSplat is a novel generalizable 3D reconstruction framework that extends the field-of-view from sparse input images by leveraging semantic priors. It introduces an Open Gaussian Growing approach that combines RGB-semantic joint inpainting guided by bidirectional diffusion models to extrapolate unseen regions beyond the input views. The inpainted images are then lifted back into 3D space for progressive optimization of Gaussian parameters, enabling efficient and high-quality semantic-aware scene reconstruction.

Results

Open-Vocabulary Query

Quantitative Results on GO Benchmark

OGGSplat effectively extrapolates to unseen regions while maintaining both high visual fidelity and semantic plausibility.

Qualitative Analysis on Real-World

OGGSplat demonstrates strong semantic-aware scene reconstruction capabilities, even when handling out-of-distribution samples or just two-view images captured directly from a smartphone camera.

Citation

Bibtex

            @article{wang2025oggsplat,
              title={OGGSplat: Open Gaussian Growing for Generalizable Reconstruction with Expanded Field-of-View},
              author={Wang, Yanbo and Wang, Ziyi and Zheng, Wenzhao and Zhou, Jie and Lu, Jiwen},
              journal={arXiv preprint arXiv:2506.05204},
              year={2025}
            }