Pointnet_Pointnet2_pytorch

Pointnet_Pointnet2_pytorch copied to clipboard

I am just starting learning PointNet2. I want to know why the data is six dimension instead of three dimension, such as "x,y,z".And if i want to train my data, what I should do to deal with my own dataset?

Hi, Dimension Issue: as per the paper dimensions can be extended based on the additional features. For 6 dimensions, it should be the x,y,z coordinate along with the vertex color. For 9 dimensions, you can have the normal values for the vertex as well. Training Issue: I have just started to use this implementation, will update this thread when I train my own dataset.

Thank u for your answer.I have already solved this problem.When I train my data,the most important is the delimiter in our data.In this project,',' is used to be a char that splits data,so we should care about our data or change the delimiter in code.

Thanks for updating. Have you tried the network on a windows machine too? Does it require cuda support when running on Linux?

Thanks for updating. Have you tried the network on a windows machine too? Does it require cuda support when running on Linux? I tried this network on Ubuntu18.04.But i think it absolutely can run in Win as long as u have already installed Pytorch.It also need CUDA.

When I run the following command: python train_classification.py --model pointnet2_cls_ssg --log_dir pointnet2_cls_ssg I get this error:

  File "C:\Users\Marsar\anaconda3\envs\PointNet\lib\site-packages\numpy\lib\_datasource.py", line 533, in open
    raise IOError("%s not found." % path)
OSError: data/modelnet40_normal_resampled/lamp\lamp_0016.txt not found.

Any idea why dataset is inconsistent?

Perhaps the pathroot in Win is splited by '/' instead of ''.Because on Linux ,root is splited by '/'.U can try to change this question. Besides,it's important that your code,_datasource.py. shoud be included in the file that contains 'data/xxxxxx'. Finally,in your root,'/lamp\lamp_0016.txt' have different delimiter ,'' and '/'.I think it may be a problem

Thanks for these inputs, one problem I see is that the modelnet40_train files has image paths of classes which are not even in the modelnet40_normal_resampled. In this dataset, there is no data for Lamp though in the modelnet40_train file paths for lamp data is mentioned. What would be the quickest way to resolve this?

Thanks for these inputs, one problem I see is that the modelnet40_train files has image paths of classes which are not even in the modelnet40_normal_resampled. In this dataset, there is no data for Lamp though in the modelnet40_train file paths for lamp data is mentioned. What would be the quickest way to resolve this?

Wll,in my memory,there are other file that contains complete data from airplane_0001.txt.Whether u run the code directly or u have changed something?

No, I haven't changed anything yet. Just trying to run the cloned directory with the dataset downloaded from the given link.

OK, u can click the file,lamp, and find that the file whether contain lamp_0016.txt or not. And u can check 'modelnet40_test' that whether have the txt,because the data have been divided into train model and test model,which use differnet data.

in the code,author first read the file,'modelnet40_train',to choose the data that the model should use.U can debug the code and see whether the name of data file have already been input to the running code.

OK, u can click the file,lamp, and find that the file whether contain lamp_0016.txt or not. And u can check 'modelnet40_test' that whether have the txt,because the data have been divided into train model and test model,which use differnet data.

Thanks for the info. I checked that a some of classes are missing from the dataset but are included in the modelnet40_test and modelnet40_train files. If I remove all instances of unavailable classes, there is no error but the program stops right after epoch 1 with a bunch of binary values(I believe being read from some text file). I have to check if I can get the correct dataset to begin with. I will need to debug it.

wish u success

Thanks. I redownloaded the dataset from the same link, and to my surprise it has all the classes! I ran the same script and now it training fine(still training). Thanks!

(I am on windows, if it helps anyone)

I realize that gpu usage by the pointnet_cls model is very less, intermittent and max upto 7%. I used Nvidia GeForce 1070TI.

I ended up reducing the epochs to 100 and stopped training at 32 epochs. Test accuracy came out to be 82% which was okay for me, as I only wanted to inspect how network is learning.

OK,congratulation !