how does extract center point from partial point cloud for application?

[jinyoung-kim-dinnercoding]

Thank you for your amazing stuff. I trained my custom data. And I have some problems. I understand that before training point cloud completion, it needs to find center point of complete point cloud. And then make partial point cloud to train.

In the terms of application, I hardly extract center point of complet point cloud, cuz I get partial data from depth camera. Here is the example.

Like above picture, center point is moved by missing point cloud.

I have these idea.

1. centering partial point cloud not complete point cloud during training.
2. augmentation of translation center point of complete point cloud.

I'm not sure it will works. I should do experiments.

Could you share your insight, please?

I found similiar issue below. https://github.com/yuxumin/PoinTr/issues/61

[yuxumin]

Hey there, @jinyoung-kim-dinnercoding!

Regarding the adoption of the point completion network into other applications, I believe that the biggest challenge lies in its heavy dependence on the input data distribution and normalization. When dealing with partial point clouds, we often need to adjust them to resemble a sample within the training dataset.

However, what if we cannot achieve this similarity? For instance, we may not know the center, which is a common scenario. In my view, idea 1 and idea 2 are essentially the same - both seek to broaden and remove constraints from the input distribution. Incorporating some prior knowledge, such as category, which can be encode as other input for the network, could prove useful in making the approach more practical.

[jinyoung-kim-dinnercoding]

Thank you for sharing your insight! I will do experiments.