search

cvg / pixloc

Back to the Feature: Learning Robust Camera Localization from Pixels to Pose (CVPR 2021)
Apache License 2.0
735 stars 92 forks source link

Using COLMAP to generate Cambridge landmark scene's poses from scratch #33

Open patelajaychh opened 2 years ago

patelajaychh commented 2 years ago

Hi, I'm trying to do SFM reconstruction of Cambridge landmarks scenes using Superpoint features and Superglue matches based database (created using HLoc tool) in COLMAP. I'm wondering, will this be correct way? I will be using colmap mapper for creating 3D model from reference images only and then will use colmap image_registrator with fix_existing_images=True to register query images to get poses information. Reference images only model will be input to PixLoc while model output after registering query images will be used as ground truth for evaluating PixLoc algorithm.

If this this not correct, then what should be ideal method to generate poses for images without using nvm file provided with dataset?

tsattler commented 2 years ago

If you want to compare your results against those published in the literature, you should use the poses in the nvm file.

patelajaychh commented 2 years ago

Thanks for quick response. I'm interested in evaluating with my own model and grouth truth.

tsattler commented 2 years ago

Why not use the poses from the nvm file then.

patelajaychh commented 2 years ago

I'm developing a generalized pipeline for my own data where I don't have pose information, only RGB images are there. So I'm testing that with Cambridge landmarks.