Detail Preserved Point Cloud Completion via Separated Feature Aggregation

Detail Preserved Point Cloud Completion via Separated Feature Aggregation ECCV 2020

Wenxiao Zhang, Qingan Yan, Chunxia Xiao

Wuhan University; JD.com American Technologies Corporation, CA

Update 2022/6/13

To test and generate the submission.zip file for completion3d leaderboard, run eval_completion3d.py. We follow the way of GRNet which directly use the pre-trained model on PCN dataset and downsample the results to 2048 points. You should also uncomment the L209 in rfa.py to perform the downsample.

Introduction

We propose a point cloud completion framework which leverages separated feature aggregation to reconstruct the known and the missing part separately. This work is based on PU-Net and PCN.

1) Pre-requisites

• Python3
• CUDA
• tensorflow
• tensorpack
• matplotlib
• lmdb
• msgpack
• pyarrow
• open3d-python

This code is built using Tensorflow 1.14 with CUDA 10.0 and tested on Ubuntu 18.04 with Python 3.6.

2)Compile Customized TF Operators

The TF operators are included under tf_ops and pc_distance, you need to compile them. Check the makefile under pc_distance folder and compile.sh under tf_ops. (You need to check the the CUDA path and tensorflow path in each tf_xxx_compile_abs.sh under each tf_ops subfolder) first. Refer to Pointnet++ for more details.

Important: We modified the original source code gather_point function in tf_sampling.cpp and tf_sampling_g.cu to fit our needs. If you have already complied the sampling operator in other project, you need to recompile it.

3)Download pre-trained models

Download pre-trained models on trained_models folder from Google Drive and put them on data/trianed_models dir. The pre-trained models consist networks of RFA and GLFA.

4) Testing

1. Download ShapeNet test data on Google Drive. Specifically, this experiment requires test.lmdb and test_novel.lmdb. Put them on data/shapenet folder.
2. Run python3 eval.py. Use --model_type option to choose different model architectures. Use --save_path option to choose the folder to save the qualitative results. If you just want to see the quantitative results, you can set --plot to False for fast evaluation.

Note that we use the same testing data in PCN project but we create the lmdb file for faster evaluation. You can also use the original pcd file provided on PCN repository.

5) Traning

1. The training data are from PCN repository, you can download training (train.lmdb, train.lmdb-lock) and validation (valid.lmdb, valid.lmdb-lock) data from shapenet or shapenet_car directory on the provided training set link in PCN repository. Note that the training data for all 8 categories in shapenet takes up 49G of disk space. The training data for only the car category takes 9G instead.
2. Run python3 train.py. Use --model_type option to choose different model architectures. Type python3 train.py -h for more options.