fpsandnoob/pointshufflenet

Pytorch Implementation of PointShuffleNet

This code is for classification testing on ModelNet40

Environments

Ubuntu 18.04 or Windows10
Python 3.6 or above
CUDA 10.2
Pytorch 1.4 or above

Package

Pytorch
torchsummary
tqdm
prefetch_generator
h5py

Data Preparation

You need to download the test data file from GoogleDrive and put it in the data directory as data/cache_test_1024_normal_True.h5. The data is generated from ModelNet and downsampled to 1024 points with normal vectors. The details of generating is in class ModelNeth5DataLoader of data_utils/ModelNetDataLoader.py.

Model

The default model in test_cls.py is models/PN2_cls_ssg_shuffle_info.py which is our PointShuffleNet with HER and LMIR. For stable evaluation result, we don't use ClusterFPS in default model. But you can try it by changing the use_cluster parameter in models/PN2_cls_ssg_shuffle_info.py.
The code of HER and LMIR is in models/ShufflePointNet_util.py line 61.
The code of ClusterFPS is in models/ClusterFPS.py

Run

python test_cls.py

Performance

You should get the result as

Test Instance Accuracy: 0.931048, Class Accuracy: 0.910952