Reproduce “TSGaussian: Semantic and Depth-Guided Target-Specific Gaussian Splatting from Sparse Views” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Liang Zhao, Zehan Bao, Yi Xie, Hong Chen, Yaohui Chen, Weifu Li
TSGaussian presents a 3D Gaussian Splatting framework for reconstructing specific target objects from sparse 360-degree views, integrating semantic and multi-scale depth constraints. The approach achieves clearer novel view synthesis and more accurate segmentation compared to baselines, showing a PSNR improvement of 6.28 on the "bear" dataset and 8.50 on a custom "Citrus" dataset over 3DGS and Gaussian Grouping, respectively, while preventing geometric degradation and semantic inconsistency.
Key insights
- Proactively integrating semantic constraints into the 3DGS training process, rather than relying on post-reconstruction segmentation, is crucial for guiding target-specific reconstruction and ensuring semantic accuracy.
- Combining both semantic guidance and multi-scale depth regularization effectively mitigates common issues like geometric degradation and artifact generation when reconstructing from sparse views.
- Restricting Gaussian densification and pruning to a Region of Interest defined by semantic attributes significantly reduces computational redundancy and focuses resources on the designated target object.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.