Radiance Fields from VGGSfM and Mast3r, and Their Comparison
Project Overview
This project is a comparative study focused on evaluating the performance of two different deep learning-based Structure-from-Motion (SfM) methodologies, VGGSfM and Mast3r, in the context of Radiance Fields reconstruction using Gaussian Splatting. The primary objective was to understand the strengths and weaknesses of each approach when applied to wild, internet-sourced image datasets.
By implementing and testing both VGGSfM and Mast3r pipelines for generating 3D point clouds and camera poses, and subsequently using these outputs for 2D Gaussian Splatting, this project provides insights into the suitability of each SfM method for inverse rendering tasks. The comparison includes qualitative visual assessments of the reconstructed point clouds and Radiance Fields, as well as a quantitative analysis of their characteristics.
Key Contributions:
- Implemented Gaussian Splatting pipelines using both VGGSfM and Mast3r for 3D reconstruction.
- Developed a script to convert MASt3R's output to a COLMAP-compatible format for easier integration with existing visualization and rendering tools.
- Conducted experiments on diverse scenes to compare the performance of VGGSfM and Mast3r.
- Analyzed and summarized the results, highlighting the advantages and limitations of each method for Radiance Fields reconstruction.
Project Details
- Category: Research, Comparative Study, 3D Reconstruction
- Project URL: Project Github
- Skills Demonstrated: Structure from Motion (SfM), Gaussian Splatting, 3D Visualization, Comparative Analysis
Key Features & Findings
Findings
Key observations from the comparative study:
- MASt3R: Provides denser and more diverse point cloud reconstructions, but camera pose accuracy is less suitable for inverse rendering.
- VGGSfM: Offers more accurate camera pose reconstruction due to Bundle Adjustment, making it better suited for inverse rendering tasks, despite sparser point clouds. Camera poses are significantly more accurate compared to MASt3R.
- Robustness: For sparse views, both VGGSfM and MASt3R are more robust than traditional COLMAP, successfully reconstructing datasets where COLMAP fails.
- VRAM Efficiency: VGGSfM demonstrates better VRAM efficiency, capable of handling larger datasets compared to MASt3R.
- Pose Refinement Potential: MASt3R and VGGSfM poses can serve as effective initializations for camera pose optimization during Radiance Field training.
Results Visualization
1) COLMAP PointCloud Comparison
MASt3R | VGGSfM |
---|---|
2) Radiance Fields Reconstruction Comparison
MASt3R | VGGSfM |
---|---|
![]() |
![]() |
3) MASt3R + Camera Pose Optimization