How to use hloc to obtain accurate poses when only using a camera to capture an indoor scene？

[Bin-ze]

Noting the excellent performance of HLOC, I wanted to use it for pose estimation in indoor scenes. I collected multiple sets of indoor data, but when visualizing poses, I found that in many cases, there were incorrect pose estimates.

I used the following configuration:

     matching_method: Literal["vocab_tree", "exhaustive", "sequential"] = "sequential",
     feature_type: Literal[
         "sift", "superpoint_aachen", "superpoint_max", "superpoint_inloc", "r2d2", "d2net-ss", "sosnet", "disk"
     ] = "superpoint_aachen",
     matcher_type: Literal[
         "superglue", "superglue-fast", "NN-superpoint", "NN-ratio", "NN-mutual", "adalam"
     ] = "superglue",

Please what kind of technique can I use HLOC to get a better pose than colmap in indoor scenes?

[aboutyy]

I had the same problem, I guess maybe it's the low texture problem.

[Bin-ze]

I want to know how to extract distinguishable features for indoor low texture?

I found that indoors, hloc has a very high failure rate. In my experiments, more than 80% of scene captures could not be used to train a detailed nerf algorithm.

This greatly increases the difficulty of indoor 3D modeling, because relying solely on image information does not seem to be enough to obtain usable poses.

If you have any suggestions I would be very grateful