The direction and angle of the point cloud will affect the inference effect

[iyinlin]

Hello, thank you for your great work!

I am a beginner in point cloud completion, and when using this code to infer other airplane point clouds in Shapenet, I found that the effect was poor, while the original 'airplane.pcd' file in the demo directory worked very well. I checked various details such as normalization and preprocessing, and there were no issues.

Later, I accidentally discovered that my point cloud and the original point cloud had a different direction. To verify that it was a problem with the direction, I also rotated the original airplane in the same direction. As shown in Figure 1, the rotated plane was placed in the same window for visualization.

Why is there such a significant difference in inference performance when only rotating the point cloud 90 degrees? How can I ensure that my point cloud direction is correct during testing to achieve good inference results? Or can random rotation be added during training to enhance robustness? (Sorry, I'm a beginner and I'm not sure if this is feasible.)

[SLDRMK]

I came into a similar situation. The net can often complete the pc with default poses greatly, while can hardly align them. I tried to train one with such RandomRotate transforms, but the final result is awful. I am not sure if this is affected by the net structure.