Reproduce “An Evaluation of DUSt3R/MASt3R/VGGT 3D Reconstruction on Photogrammetric Aerial Blocks” 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
Xinyi Wu, Steven Landgraf, Markus Ulrich, Rongjun Qin
An evaluation of transformer-based 3D reconstruction models (DUSt3R, MASt3R, VGGT) against traditional photogrammetry (COLMAP) on aerial imagery reveals their robustness in low-overlap conditions and superior processing speed. This study demonstrates that learning-based approaches offer valuable completeness for sparse data and mesh completion, complementing the higher accuracy of traditional methods in standard scenarios.
Key insights
- Learning-based methods, particularly VGGT and MASt3R, demonstrate robustness in generating point clouds and recovering poses from extremely sparse image sets or very low overlaps, scenarios where traditional photogrammetry often fails.
- COLMAP, when provided with high-resolution imagery and generous overlap, consistently yields superior absolute accuracy in dense point cloud generation and camera pose estimation.
- Transformer-based models exhibit a drastic speed advantage over traditional methods, with VGGT processing images significantly faster (e.g., 1% of COLMAP_HR's time for 38 images).
- Learning-based models serve as valuable complementary tools rather than direct replacements for traditional photogrammetry, particularly for tasks like mesh completion in areas with sparse coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.