Is the performance realtime?

[nonlinear1]

Your code is very good. I am coding a slam with 3D descriptor, which is trained by CNN. The 3DMatching of your codes with TSDF to TSDF registration can run in realtime?

[wangziyang182]

It is hard for it to run in real time. It need some modifications for that to happen. The structured 3D tensor plus the way I handling the matching point make it kind of slow. To achieve real time, dense voxel wise inference is very expensive. You can try to train instead of contrastive loss, using classification loss which will make it much faster. However, it might sacrifice the accuracy to some extent.

[nonlinear1]

Thank you very much! Do you have wechat(微信)？If you like, I can add your wechat?