HengyiWang / Co-SLAM

[CVPR'23] Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM
https://hengyiwang.github.io/projects/CoSLAM.html
Apache License 2.0
399 stars 35 forks source link

reconstruction #53

Closed mohcenaouadj closed 1 month ago

HengyiWang commented 1 month ago

Hi @mohcenaouadj,

That is a great question. In fact, since you already know that you only have one object around the center, you can do a good initialization of the SDF field. In our new paper MorpheuS, we initialize the SDF into an approximate sphere using something similar to https://arxiv.org/abs/2002.10099. I think you can try this implicit initialization or just use a fixed blob function and let the model fit the SDF residual.

mohcenaouadj commented 1 month ago

@HengyiWang thank you a lot for your answer

HengyiWang commented 1 month ago

This is because the coordinate encoding is continuous. With the inherent smoothness of MLPs, once you start to train the model, it can have some sort of extrapolation (It naturally fits indoor scene reconstruction as most parts of the scene are flat (floor, wall, etc).

mohcenaouadj commented 1 month ago

Thanks a lot