DSOPP was designed as a modular system. Almost every component could be replaced at runtime, providing all needed classes and creation from yaml config in factory files.
This helps to use this code for research purposes.
For example, if you came up with a new Rolling Shutter formulation, then all you need is to create another Motion class in src/energy/motion. And create dsopp<Motion> class in src/application/dsopp_main.cpp with your own brand new motion model. Then try to compile, and implement new methods like derivatives in reprojector/projector files. Then try to compile, implement some more methods, and it is done.
To implement new camera model in dsopp you need to create model class in src/energy/camera_model. Construct of your model in src/sensors/camera_calibration/src/fabric.cpp from calib.txt file.
Let us imagine that you have trained NN to predict transformation from one image into another and you want to use it instead of a frontend solver.
You'll need to create new pose_aligner in src/energy/problems/include/energy/problems/pose_alignment/ and construct it from mono.yaml in src/tracker/tracker/src/fabric.cpp inside createPoseAlignment function.
Want to use some image embedding to spice the backend solver up? No problem! Implement embedder code somewhere. Add desired instantations on the bootom of src/energy/problems/src/eigen_pose_alignment.cpp, pass your new images to solver in src/tracker/tracker/src/monocular_tracker.cpp in estimatePose function.