search

cvg / Hierarchical-Localization

Visual localization made easy with hloc
Apache License 2.0
3.2k stars 593 forks source link

Feature and match files for detection-free algorithms #59

Closed ufukefe closed 3 years ago

ufukefe commented 3 years ago

Hi, thanks for your great work!

I want to ask that how I should organize feature and match files for algorithms like NCNet and Patch2Pix which do not have external detection step.

Excuse me if I am wrong (I do not use the hloc yet, but I am planning to do), I guess I should prepare these files like this way:

For InLoc dataset, for example, let's say I have a query image Q1 and its corresponding NetVLAD pairs {D1,D2,...,D40} from database images, and I have different matches for every pair Q1-D1, Q1-D2... these are outputs of the above-mentioned algorithms.

So, should I arrange the feature file for Q1 as the union of all matched points in the query image for each corresponding database image? After that, should I prepare the match file accordingly?

sarlinpe commented 3 years ago

Hi @ufukefe, It is not trivial to do SfM 3D reconstruction with dense matching, because there is no notion of track. I am not aware of any right way of doing this. It seems that Patch2Pix groups neighboring pixels together into a single track. This is explained in Section C of the paper. The authors have a fork of hloc here, but I haven't tried it out.