CuiRuikai
commented
2 months ago I cannot understand what you are trying to achieve, so I will explain my understanding, and please correct me: Suppose you have a square surface region to complete, the input is the centre part (as you said the expected completion model located at the four corners), but the inference result is only points in the centre part.
I guess this might be an issue related to the way that we create local regions. We use farthest point sampling to sample n patch centres. Then, for each centre, we collect the k nearest neighbours. If you visualise KNN neighbourhood, they tend to be a sphere-like shape instead of a smooth surface. This is acceptable for objects as there are natural curvatures and complete topology, but for a smooth surface, this might cause some distortion to the learning. However, I am not sure how this is related to your problem.
Scale is also an important factor in the training. Although we did not normalize objects in the dataloader, the data itself already normlized. ShapeNet derived data is usually normalized to [-1, 1]. If your data is unnormalized, then our method won't work.
Dear author,
Thank you for addressing everyone's questions! When I use a self-made dataset (smooth metal surface) to train a model for completion, the completion effect is significantly different from the expected result. The result is a bunch of discrete points in the middle, with the expected completion model located at the four corners. However, the desired effect is only a single expected completion model. Is it an issue with the dataset or is the model structure not suitable for the dataset?
Look forward to your reply！