Inference results on my own dataset

[ranhaogong]

I have trained a new model on my own dataset with normlization applied.
Then I checked the inference result, but found bed results.

the input partial point cloud is as followed:

the complete point cloud is as followed:

the predicted result (with out merging with input partial):

the merged result:

I wonder why and looking forward to your suggestion!

[CuiRuikai]

Our method is designed for object. Adopting to scene level completion does lead to a lot of problems. This is aligned with my expectation. The way we create patch needs knn and it is likely the reason. See a recent paper [1] from ECCV 2024 shows similar problem of our method in Fig 3. Our method is not good at producing large planes. It is prone to create spherical clusters.
let me know if you have further concerns
[1] https://arxiv.org/pdf/2409.15939