search

likojack / bnv_fusion

This repository implements our CVPR2022 paper "BNV-Fusion: Dense 3D Reconstruction using Bi-level Neural Volume Fusion"
MIT License
135 stars 18 forks source link

the ground truth mesh for 3dscene seems cut in come cube #14

Closed Jarrome closed 1 year ago

Jarrome commented 1 year ago

I compared with ScanNet that looks fair.

But for 3Dscene (just find, not check ICL yet), the mesh seems trimed. See the sharp cut line on the floor.

Screenshot from 2023-03-13 21-56-50

While mine result has a lot more, Screenshot from 2023-03-13 21-59-35

This cause around 68 acc which is quite close to DI-Fusion.

So I'm wondering if author also cut the DI-Fusion? Or DI-Fusion also use the same un-cutted for compare?

Otherwise, how can I cut my mesh to have a fair compare with bnv-fusion?

likojack commented 1 year ago

The GT mesh provided by the dataset is cut for some reason. We have to truncate the depth maps to 3m (see here) to match the GT mesh.

Jarrome commented 1 year ago

Emmm, I also use depth map truncated to 3m... This does not explain the clean cut because you truncated on local frame...

I guess maybe because you use bound [-1,1]^3 ?

likojack commented 1 year ago

We truncated the extracted mesh based on the dimensions of the GT mesh, see here

Jarrome commented 1 year ago

Ah, ok. Thanks for the help. :D