ARFNet

The codes for Adaptive Recurrent Forward Network for Dense Point Cloud Completion published by TMM. This work is the enhanced version of RFNet published in ICCV2021. It works by replacing the parameter-controlled merge layer in RFNet with network-controlled Adamerge module. Refine Cell in RFNet is not used in this work for efficiency, which can also be added to further improve the performances.

Environment

• TensorFlow 1.13.1
• Cuda 10.0
• Python 3.6.9
• lmdb 0.98
• tensorpack 0.10.1
• numpy 1.14.5

Dataset

The adopted dataset can be found in PCN.

Usage

1. Compile

cd ./tf_ops
bash compile.sh

2. Train

Python3 vv_recon.py

Note that the paths of training data(trainpath) and validation data(valpath) should be edited according to your setting.

3. Test

Python3 recon_test.py

The paths of test data(data_dir) and lists(list_path) should be edited before testing. The qualitative results should be

The quantitative results on the Known categories of ShapeNet in PCN would be

Citation

If you find our work useful for your research, please cite:

@article{huang2022adaptive,
  title={Adaptive Recurrent Forward Network for Dense Point Cloud Completion},
  author={Huang, Tianxin and Zou, Hao and Cui, Jinhao and Zhang, Jiangning and Yang, Xuemeng and Li, Lin and Liu, Yong},
  journal={IEEE Transactions on Multimedia},
  year={2022},
  publisher={IEEE}
}