Improving incremental Structure-from-Motion

Contributions of this paper marked with 🔴

An architecture of incremental Structure-from-Motion

Correspondence Search

• input : set of images
• output : set of geometrically verified image pairs and a graph of image projections for each point

1. Feature Extraction

• SIFT detects sets of local features at locations for each image

2. Matching

• SfM discovers images that see the same scene part by leveraging the features $F_i$ as an appearance description of the images

3. Geometric Verification

• SfM verifies the matches by trying to estimate a transformation that maps feature points between images using projective geometry.

🔴 Scene Graph Augmentation

• Augmenting the scene graph with the appropriate geometric relation
  • An already augmented scene graph enables to efficiently find the optimal initialization for a robust reconstruction process

Incremental Reconstruction

1. Initialization

• SfM initializes the model with a carefully selected two-view reconstruction