Skip to content

Multi-view stereo helpers (pytheia.mvs)

pyTheia exposes a small MVS-oriented API from the C++ Theia tree. Bindings: mvs.cc.

ViewSelectionMVSNet

pt.mvs.ViewSelectionMVSNet(reconstruction, num_neighbors, theta0=5.0, sigma1=1.0, sigma2=10.0)

Selects, for each view, a ranked list of neighbor views suited for MVSNet-style depth inference, using covisibility and triangulation angle cues as in:

Yao et al., MVSNet: Depth Inference for Unstructured Multi-view Stereo, ECCV 2018.

Returns: dict[int, dict[float, int]] — outer key is a view id; each inner dict maps a score (float) to a neighbor view id (int), with scores ordered for best-first iteration (C++ std::map<double, ViewId, std::greater<double>>).

Parameters:

  • num_neighbors — how many neighbors to retain per view (interpreted with the scoring function in C++).
  • theta0, sigma1, sigma2 — angular / weighting parameters (see view_selection_mvsnet.h and the .cc implementation).

This does not run a learned MVS network; it only proposes view subsets given an existing Reconstruction with cameras and tracks.

See also

  • SfM — building a Reconstruction before view selection
  • Triangulation — structure needed for meaningful covisibility