PCDNet

[ECCV'22] The codes for Resolution-free Point Cloud Sampling Network with Data Distillation

Environment

• TensorFlow 1.13.1
• Cuda 10.0
• Python 3.6.9
• numpy 1.14.5

Dataset

The adopted ShapeNet Part dataset is adopted following FoldingNet, while the ModelNet10 and ModelNet40 datasets follow PointNet. Other datasets can also be used. Just revise the path by the (--filepath) parameter when training or evaluating the networks. The files in (--filepath) should be organized as

    <filepath>
    ├── <trainfile1>.h5
    ├── <trainfile2>.h5
    ├── ...
    ├── train_files.txt
    └── test_files.txt

where the contents in (train_files.txt) or (test_files.txt) should include the directory of training or testing h5 files, such as:

    train_files.txt
    ├── <trainfile1>.h5
    ├── <trainfile2>.h5
    ├── ...

Usage

1. Preparation

cd ./tf_ops
bash compile.sh

The pre-trained reconstruction and recognition networks are conducted following S-Net. They should be included in the parameter (--prepath).

2. Training

For the reconstruction task,

Python3 pc_sampling_rec.py

For the recognition task,

Python3 pc_sampling_cls.py

Note that the path of data (--filepath) should be edited according to your setting.

3. Evaluation

For the reconstruction task,

Python3 eva_rec.py

For the recognition task,

Python3 eva_cls.py

The trained weight files should be put in (--savepath) to evaluate the sampling performances.