Philipp Lindenberger

Add semi-dense matching support (LoFTR)

2

- LoFTR module (`hloc/matchers/loftr.py`) - semi-dense matching on image pairs (`hloc/match_dense.py`) - quantization to store features and matches in hloc format - supports quantization to pre-extracted features (e.g. superpoint)

- Allow extracting costmaps from dense features - Store features to H5 in dense if it requires less memory (e.g. for semi-dense local features)

Releases GPU resources before/after dense feature extraction (by moving the model to CPU).

Add support for [kornia's DenseSIFTDescriptor](https://kornia.readthedocs.io/en/latest/_modules/kornia/feature/siftdesc.html#DenseSIFTDescriptor) as dense features. Requires kornia>=0.6.4.

Updates the config of the sift extractor in the ETH3D benchmark to be compatible with the latest API changes in hloc.

- eval [LoFTR](https://zju3dv.github.io/loftr/) features+matches on ETH3D (recommended: use `low_memory` conf) with [hloc interface](https://github.com/cvg/Hierarchical-Localization/blob/master/hloc/match_dense.py).

Problem: COLMAP uses [Eigens aligned allocator](https://eigen.tuxfamily.org/dox/group__TopicStlContainers.html) for vectors of [Point2D](https://github.com/colmap/colmap/blob/dev/src/base/point2d.h#L96). This PR ensures that `pycolmap.ListPoint2D` also uses this aligned allocator (previously caused `MaxRecursionError` in some cases).

Reorganize the readme by highlighting the core reconstruction API on top and describe how to retrieve available and default parameters.

- Add poisson and delaunay meshing (#92) - Delaunay meshing is only supported if COLMAP was compiled with CGAL support

Hit the subscribe button on the right of this issue if you wish to be notified of the training and evaluation code release in a separate repo. Please do not...