NoteNextra-origin/content/CSE5519/CSE5519_G3.md

CSE5519 Advances in Computer Vision (Topic G: 2023: Correspondence Estimation and Structure from Motion)

Detector-Free Structure from Motion

link to paper

• A new detector-free SfM framework built upon detector-free matchers to handle texture-pool scenes.
• An iterative refinement pipeline with a transformer-based multi-view matching network to efficiently refine both feature tracks and reconstruction results.
  • Multi-view Feature transformer to enhance the discrimitiveness of extracted features.
  • Use Bundle adjustment (view point consistency)
  • Use Topology adjustment (merge, complete, or remove vertices using pre defined rules)

Tip

This paper proposed a new detector-free SfM framework built upon detector-free matchers to handle texture-pool scenes and use an iterative refinement pipeline with a transformer-based multi-view matching network to efficiently refine both feature tracks and reconstruction results.

I'm particularly interested in the detector-free matchers and the transformer-based multi-view matching network. Due to time constraints, I don't have much time to check the work for detector-free matchers and how they generate the coarse model for predicted matches. I'm looking forward to hearing more about this topic in tomorrow's presentation.