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

How to add new query image without reconstruct hloc db? #28

Open Arlen0615 opened 2 years ago

Arlen0615 commented 2 years ago

Hi, Thanks for great work and it's really useful to remapping scene.

Please allow me simplify the own data process first:

  1. Prepare images of scene
  2. Run hloc to get A. db of sfm, B. matches keras model, C.pairs_info.txt(e.q. pairs-exhaustive.txt, pairs-query-netvlad50.txt)
  3. Run pixloc to localize query image

I check the code in localizer.py, it need get the id from pairs_info.txt and sfm db self.model3d = Model3D(paths.reference_sfm) dbs = [self.model3d.name2id[r] for r in self.retrieval[name]]

Is that mean I have to do step 2 every time once new query image added?

lck666666 commented 2 years ago

I have the same question, if so, this is not a real-time inference.

famfa693 commented 2 years ago

Thanks for your amazing work! I have the same question. Anyone have any thought on this? Is this a real-time 6 DOF pose estimation? What is the minimum inference time we can get from this approach? Thanks in advance.

tieguanyin803 commented 1 year ago

I have the same question

tsattler commented 1 year ago

@Arlen0615 It is not necessary to do all parts of step 2 to localize a new image. You only need to determine which database images are relevant for the new image.

Regarding timings: No, this is not a real-time pose estimation algorithm as it is computationally quite expensive.