This code is for classification testing on ModelNet40
Ubuntu 18.04 or Windows10
Python 3.6 or above
CUDA 10.2
Pytorch 1.4 or above
Pytorch
torchsummary
tqdm
prefetch_generator
h5py
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
.
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
python test_cls.py
You should get the result as
Test Instance Accuracy: 0.931048, Class Accuracy: 0.910952