search

HengyiWang / spann3r

3D Reconstruction with Spatial Memory
https://hengyiwang.github.io/projects/spanner
421 stars 16 forks source link

Tips for large scale scene reconstruction #2

Open LaFeuilleMorte opened 1 week ago

LaFeuilleMorte commented 1 week ago

Hi, Thanks for your brilliant work! I was trying to reconstruct a whole indoor scene with this model BUT the results seemed not good. So I decreased the kf_every param 10 -> 1, The point cloud becomes denser but still not good enough. Is there any tips on reconstructing more than 300 images? I tried dust3r and it can't reconstruct so many views. I hope this model could solve my problems. The point cloud and one of the image collections were like the following images:

image

image

HengyiWang commented 1 week ago

Hi @LaFeuilleMorte,

If by "not good" you mean some parts of the scene are missing, you might want to try adjusting the --conf_thresh. In terms of the reconstruction quality, the current version of Spann3R is not as good as DUSt3R since it is a real-time reconstruction pipeline. If you need a super high-quality scene reconstruction for your application without the requirements for running it online in real-time, you can try some optimization-based approach (e.g. 2DGS).