LFNet: Local Rotation Invariant Coordinate Frame for Robust Point Cloud Analysis
Install TensorFlow. The code is tested under TF1.13 GPU version and Python 3.6 on Ubuntu 16.04.
The TF operators are included under tf_ops, you need to compile them (check tf_xxx_compile.sh under each ops subfolder) first. Update nvcc and python path if necessary. The code is tested under TF1.13. If you are using earlier version it's possible that you need to remove the -D_GLIBCXX_USE_CXX11_ABI=0 flag in g++ command in order to compile correctly.
To compile the operators in TF version >=1.4, you need to modify the compile scripts slightly.
First, find Tensorflow include and library paths.
TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
Then, add flags of -I$TF_INC/external/nsync/public -L$TF_LIB -ltensorflow_framework to the g++ commands.
Shape Classification
Download and unzip ModelNet40 (1.6G). Replace $data_path$ in cfgs/config_*_cls.yaml with the dataset parent path.
ShapeNet Part Segmentation
Download and unzip ShapeNet Part (674M). Replace $data_path$ in cfgs/config_*_partseg.yaml with the dataset path.
To train a PointNet++ model to classify ModelNet40 shapes (using point clouds with XYZ coordinates):
python train.py
After training, to evaluate the classification accuracies (with optional multi-angle voting):
python evaluate.py
You can use our model cls/model_iter_113_acc_0.905592_category.ckpt as the checkpoint in evaluate.py, and after this voting you will get an accuracy of 91.08% if all things go right.(90.56% during training)
To train a model to segment object parts for ShapeNet models:
cd shapenet_seg
python train_seg.py
evaluate:
cd shapenet_seg
python evaluate_shapenet.py
You can use our model seg/model_best_acc_inst.ckpt as the checkpoint in evaluate_shapenet.py, and after this voting you will get instance mIoU of 81.13% if all things go right.(81.01% during training)
The code is heavily borrowed from A-CNN.
If you have some ideas or questions about our research to share with us, please contact [email protected]