search

charlesq34 / pointnet

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
Other
4.71k stars 1.44k forks source link

Data augmentation for part segmentation? #140

Open ThibaultGROUEIX opened 5 years ago

ThibaultGROUEIX commented 5 years ago

Hi, awesome work and great repo ! I am trying to reimplement it in pytorch and get the paper's result of 83.7 mean IOU for part segmentation.

To do so, I have two questions on part segmentation : 1)- Do you perform data augmentation for part segmentation on Shapenet16 ? I couldn't find the info in the paper or the code.

If yes: could you say which kind ? If no: why do it for classification on modelnet40 and not on shapenet16 ?

2)- What is the difference in performance with and without the T-NET ?

Thanks in advance for your help, Best regards, Thibault

ThibaultGROUEIX commented 5 years ago

It seems that you do not do data-augmentation indeed, from your calls to provider.py in train.py. Do you have insight on why ? Can you confirm that no further preprocessing than normalization is done ?

guobaisong commented 5 years ago

It seems that you do not do data-augmentation indeed, from your calls to provider.py in train.py. Do you have insight on why ? Can you confirm that no further preprocessing than normalization is done ?

Hi, awesome work and great repo ! I am trying to reimplement it in pytorch and get the paper's result of 83.7 mean IOU for part segmentation.

To do so, I have two questions on part segmentation : 1)- Do you perform data augmentation for part segmentation on Shapenet16 ? I couldn't find the info in the paper or the code.

If yes: could you say which kind ? If no: why do it for classification on modelnet40 and not on shapenet16 ?

2)- What is the difference in performance with and without the T-NET ?

Thanks in advance for your help, Best regards, Thibault

Hi, can you get the accuracy in paper without data augment?