Object Level Experiment on PointNet, ModelNet 10

PointNet, Task: cls

ModelNet 10， original validation acc:

After adding gaussian noise(mean 0, std 0.02), accuracy dropped to 80%

Then, after unadv perturbation, accuracy improved to 98%

unadv perturb <-> Denoising?

Original Valid accuracy: 80 % Perturbed Valid accuracy: 98 %