Hi, thanks for open sourcing this work
I have tried the demo of point cloud conditioning and it works well.
However, when I tried my use cases, I found it has two limitations (failure cases)
1. the input point cloud is incomplete due to occlusion
2. the point cloud has some outliers
I think it's becasue of the training strategy. As claimed in the paper, input point cloud covers 3 types: from depth maps / scans / complete point clouds, but the occlusion is NOT considered. Perhaps it can be migitated by randomly dropping out some points during training
Hi, thanks for open sourcing this work
I have tried the demo of point cloud conditioning and it works well.
However, when I tried my use cases, I found it has two limitations (failure cases)
1. the input point cloud is incomplete due to occlusion
2. the point cloud has some outliers
I think it's becasue of the training strategy. As claimed in the paper, input point cloud covers 3 types: from depth maps / scans / complete point clouds, but the occlusion is NOT considered. Perhaps it can be migitated by randomly dropping out some points during training