# MVPGS: Excavating Multi-view Priors for Gaussian Splatting from Sparse Input Views

**Authors:** Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, Ronggang Wang  
**arXiv:** [2409.14316](https://arxiv.org/abs/2409.14316)  
**Date:** 2024-09-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.14316v1/

## Key points

- Authors address the challenge of few-shot Novel View Synthesis (NVS) in 3D vision applications
- Existing methods like Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) have limitations such as time-consuming training processes and overfitting issues
- Proposed solution MVPGS leverages multi-view priors based on 3D Gaussian Splatting to enhance geometric initialization for 3DGS
- Introduces forward-warping method with appearance constraints to prevent overfitting and improve optimization convergence
- Incorporates monocular depth regularization technique to compensate for depth estimation discrepancies
- MVPGS achieves state-of-the-art performance in few-shot NVS while maintaining real-time rendering speeds

## Summary

In their paper titled "MVPGS: Excavating Multi-view Priors for Gaussian Splatting from Sparse Input Views," authors Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang address the challenge of few-shot Novel View Synthesis (NVS) in 3D vision applications. They highlight the limitations of existing methods such as Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS), which struggle with time-consuming training processes and overfitting issues. To overcome these challenges, the authors propose MVPGS, a novel approach that leverages multi-view priors based on 3D Gaussian Splatting. By incorporating recent advancements in learning-based Multi-view Stereo (MVS), they enhance the quality of geometric initialization for 3DGS. Additionally, to prevent overfitting, they introduce a forward-warping method that adds appearance constraints based on computed geometry. This helps improve optimization convergence and ensures view-consistent geometry constraints for Gaussian parameters. Furthermore, the authors introduce a monocular depth regularization technique to compensate for any discrepancies in depth estimation. Through a series of experiments, they demonstrate that MVPGS achieves state-of-the-art performance in few-shot NVS while maintaining real-time rendering speeds. Their findings are accepted by ECCV 2024 and can be explored further on their project page at https://zezeaaa.github.io/projects/MVPGS/. Overall, MVPGS represents a significant advancement in addressing the challenges of sparse input views in NVS applications by effectively utilizing multi-view priors and incorporating innovative techniques to enhance rendering quality and efficiency.