search

zju3dv / DetectorFreeSfM

Code for "Detector-Free Structure from Motion", CVPR 2024
Apache License 2.0
612 stars 39 forks source link

neuralsfm parameters #34

Open bhack opened 6 months ago

bhack commented 6 months ago

Can you explain in the Readme some of the conf parameters? Expecially the controller section and what generally you need to adapt to test it on a new dataset as I have seen some difference between the configs in the repo.

bhack commented 6 months ago

It will also interesting in the Readme how to feed use_prior_intrin

hxy-123 commented 6 months ago

Can you explain in the Readme some of the conf parameters? Expecially the controller section and what generally you need to adapt to test it on a new dataset as I have seen some difference between the configs in the repo.

Sorry for the late reply due to the busy competition IMC24 deadline. In config, the method section generally controls which coarse matcher to use (neuralsfm.NEUSFM_coarse_matcher), coarse matching threshold (neuralsfm.NEUSFM_coarse_match_thr), etc. The controller section means whether you want to re-run a subsection of the method for debugging. For example, if you want to improve or modify the coarse SfM phase, you can skip repetitively running the coarse matching phase by setting neuralsfm.redo_matching=False and neuralsfm.redo_sfm=True. The colmap config section means control the parameters in coarse SfM pahse.

hxy-123 commented 6 months ago

It will also interesting in the Readme how to feed use_prior_intrin

use_prior_intrin parameter means feeding known intrinsic into the system, which is fixed during SfM and refinement. We use this setting in the Texturepoor SfM experiment, which you can see as a reference for how to feed known intrinsics.