By: Max Midwinter (2022)

With Credit To: Mihai Dusmanu, Johannes L. Schoenberger and True Price

Located in the sfm_free directory AKA the working directory

Create a folder called localization and git clone https://github.com/cviss-lab/localization to the folder

Run loftr_labelgen.py (change the data directory and project name to match)...

The concept is that for a scene we project 2d points to 3d then project the 3d points to the second image. We verify that the image coordinate is valid in the second image...

• Install Colmap 3.6+
  • By default, pip will install 3.4 which will not work...
    • Unless compiled Colmap from source there will be no CUDA. GPU is required for Step 6.5 (match_match_stereo) which generates depth and normal maps. 6.5 can take a very long time but can be skipped if depth maps are dense and of good quality.
  • Install Instructions: https://colmap.github.io/install.html
    • If you have difficulty installing ceres solver: https://brucknem.github.io/posts/install-ceres-solver/
    • If using older ceres solver consider: colmap/colmap#1451

• Get ROS bag from shared drive
• Launch on VM with dedicated ISO

• Depending on the raw data format you may need some misc. functions to convert your data. You can find these in util.py

Real-sense RGB-D SLAM stereo reconstructed depth maps are not working. Testing HL2 dataset