search

yanx27 / PointASNL

PointASNL: Robust Point Clouds Processing using Nonlocal Neural Networks with Adaptive Sampling （CVPR 2020）
MIT License
260 stars 34 forks source link

strange result when testing ModelNet40 with noise #20

Open amiltonwong opened 4 years ago

amiltonwong commented 4 years ago

Hi, @yanx27 ,

After training the model for ModelNet40 with --AS option, I got the following testing result for ModelNet40 with noise: log.txt.

It's strange that Evaluation with 1 Noisy Points only gives accuracy=0.040519, while the other situations give reasonable accuracy:

Noise    Accuracy
 000       0.927
 001       0.041
 010       0.904
 050       0.784
 100       0.666

Any comments for this issue?

THX!

yanx27 commented 4 years ago

@amiltonwong Hi, thank you for your attention. Do you use the same environment as I do (tensorflow version or CUDA)? I notice that your classification accuracy rate is only 92.7, which is lower than general ~93% I trained. Thus, I don't have a good solution to your problem now, could you train a new model repeatedly to double-check the results? I'll also take the time to check the code and I'm sorry for what happened to you.

amiltonwong commented 4 years ago

@yanx27 , thanks for your reply. I'll train it again to see whether the issue occurs.