PCP Model instead of PointNet

[saran-rajendran]

According to the paper, Pointnets (2017) are used for the 3D feature extraction.
However, here a modified PCP model is used.
Could you explain why it is so and why not pointnets?

[jayliu0313]

This is a good question, and there are several reasons for it:

1. Directly using the features from PointNet may not be ideal because the perceptual field is wide, making it challenging to precisely identify anomalous regions. However, by employing an local implicit network, we can precisely determine the distance of each point from the normal surface, allowing for more precise anomaly detection.
2. Compared to only using PointNet, utilizing an implicit network allows us to learn more robust geometric representations.
3. While one might intuitively consider employing a 3D network with stronger representations (such as PointNet++ or others), utilizing PointNet as a feature extractor allows for faster defect detection.

[saran-rajendran]

Hi again :)
I have tried to replace the pointnet with pointnet++. However, the scores are less compared to basic pointnet model.
Intuitively, the SDF model with Pointnet++ backbone should give better performance right?
Have you tried this approach yourself? Any idea why this behaviour?