# CSE5519 Advances in Computer Vision (Topic G: 2024: Correspondence Estimation and Structure from Motion)

## Global Structure from Motion Revisited

[link to paper](https://arxiv.org/pdf/2407.20219v1)

### Novelty in Global Structure from Motion Revisited

1. Start with Quality matches
   - Use only geometrically verified
2. Match verification strategy
   - Homography
   - Essential Matrix
   - Fundamental Matrix
3. Filtering Bad Matches
   - Cheirality test: Remove points behind the camera
   - Epipolar proximity: Remove the matches near the epipole (unstable)
   - Triangulation angle: Remove matches with small viewing angles (pool estimation for depth)
4. Track Assembly
   - Concatenate remaining matches across all image pairs
   - Form continuous tracks of the same 3D point visible in multiple views

> [!TIP]
>
> Compared with the COLMAP, the Global Structure from Motion Revisited is more robust to the noise and the outliers but less robust to repeated patterns. I wonder how this problem is resolved in normal COLMAP pipeline.