yunxiaoshi/pointnet-pytorch

PointNet in PyTorch with comprehensive experiments

pointnet-pytorch's Issues

missing txt file s3d_cat2num.txt

how can I get this text file ?

Can you give me an example if I have to make it myself?

file modelnet_cat2num.txt missing

Hi, @kentsyx ,

Would you provide file modelnet_cat2num.txt, as it is required in datasets.py?

THX!

Unable to visualize the results from the checkpoint

I am trying to visualize the results of ShapeNet part Segmentation

I followed the mentioned steps of downloading the data, unzipping it, and then ran the following command for training a model :
python train_seg.py -dset shapenet16 -r shapenet_root_dir -np number_of_points_to_sample

The training was successful :

epoch 100: 101/106 | train loss: 0.139513 | train acc: 0.949703 | train iou: 0.904224
epoch 100: 102/106 | train loss: 0.105664 | train acc: 0.961766 | train iou: 0.926347
epoch 100: 103/106 | train loss: 0.129393 | train acc: 0.950281 | train iou: 0.905272
epoch 100: 104/106 | train loss: 0.126673 | train acc: 0.950328 | train iou: 0.905357
epoch 100: 105/106 | train loss: 0.129589 | train acc: 0.953500 | train iou: 0.911132
epoch 100: 106/106 | train loss: 0.162046 | train acc: 0.321937 | train iou: 0.191851
epoch 100 | mean train acc: 0.942338 | mean train IoU: 0.894963
epoch 100 | mean test acc: 0.927583 | mean test IoU: 0.865400

In order to visualize the results, when I run the show_seg.py after doing sh build.sh I get an error like below :

model 0/375
('number of classes', 4)
Traceback (most recent call last):
  File "show_seg.py", line 46, in <module>
    classifier.load_state_dict(torch.load(opt.model))
  File "/home/arun/anaconda3/envs/deep_learning2.7/lib/python2.7/site-packages/torch/serialization.py", line 381, in load
    f = open(f, 'rb')
IOError: [Errno 2] No such file or directory: ''

It seems like the checkpoint could not be loaded.

Mistake in code

Line#45 in show_seg.py referes to PointNetDenseCls in code. But there is nothing by this name defined anywhere in the code.

mistakes in code

Hi, @kentsyx ,

I found that line96 in pointnet.py should be modified as:
return F.log_softmax(x, dim=1), trans

according to this reference.

Problem in dataset

I think there is a problem within the dataset code for S3D.
The codes read one object in the Annotations folder as a sampling unit, which is wrong. In Annotations, each object is a fraction of a scene. For example, the object can be a widow while the scene is an office. Now during training, the codes sample batch_size number of object, rather than a scene. As a result, each sample's ground truth is just an array of the same numbers. E.g., if we sampled a window object and the label of a window is 7, the ground truth is an N-dimensional vector full of 7 where N is the number of points sampled.
I think the codes should sample scenes, not object.
Did I misunderstand something? Please let me know.