Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024)

Shujuan Li* · Junsheng Zhou* · Baorui Ma · Yu-Shen Liu · Zhizhong Han

(* Equal Contribution)

Paper | Project Page

Requirements

Install python dependencies:

conda create -n apu-ldi python=3.7.12
conda activate apu-ldi
pip install torch==1.7.1+cu110 -f https://download.pytorch.org/whl/torch_stable.html
pip install open3d==0.17.0
pip install numpy einops scikit-learn tqdm h5py matplotlib pyhocon scipy trimesh

Install the built-in libraries:

cd local_distance_indicator/models/Chamfer3D
python setup.py install
cd ../pointops
python setup.py install

Metric calculation (Optional)

We use the same metrics as Grad-PU. The CGAL library and virtual environment of PU-GCN are required. And you also need to compile folder "global_field/evaluation_code" with the following commad:

cd global_field/evaluation_code
bash compile.sh

Data and Weights

Data Preparation

You need to download the PU-GAN dataset (trainset, test mesh) and PU1K dataset, and unzip them into "local_distance_indicator/data/". The structure of data folder is similar to Grad-PU, and the detail is following:

data  
├───PU-GAN
│   ├───test # test mesh file
│   ├───test_pointcloud # generated test point cloud file
│   │     ├───input_2048_16X
│   │     ├───input_2048_4X
│   │     ├───input_2048_4X_noise_0.01
│   │     ...
│   └───train
│   │     └───PUGAN_poisson_256_poisson_1024.h5
└───PU1K
│   ├───test
│   │     ├───input_1024
│   │     ├───input_2048
│   │     ...
│   └───train
│   │     └───pu1k_poisson_256_poisson_1024_pc_2500_patch50_addpugan.h5 
└───KITTI
└───ScanObjectNN

You can download the preprocessed test point clouds here. Alternertively, you can follow the instructions in Grad-PU to process your own data.

Weights

You can download pretrained models and test results .

File Organization

The final file organization as follows.

APU-LDI
├───global_field
│   ├───pretrained_global
│   │     ├───kitti
│   │     ├───pu1k
│   │      ...
├───local_distance_indicator
│   ├───pretrained_local
│   │     ├───pu1k_local
│   │     ├───pugan_local
│   ├───data
│   │     ├───PU-GAN
│   │     ├───PU1k
    ...

Test

To test APU-LDI, you can use the following command:

cd APU-LDI/global_field
# PU-GAN
python train_global.py --conf confs/pugan.conf --mode upsample --dir pugan --dataset pugan --listname pugan.txt

# PU1K
python train_global.py --conf confs/pu1k.conf --mode upsample --dir pu1k --dataset pu1k --listname pu1k.txt

# KITTI
python train_global.py --conf confs/kitti.conf --mode upsample --dir kitti --dataname scene1

# ScanObjectNN
python train_global.py --conf confs/scanobjectnn.conf --mode upsample --dir scanobjectnn --dataname chair186

For arbitrary-scale upsampling, you can modify the parameters in the config file, for example:

# data_dir = ../local_distance_indicator/data/PU-GAN/test_pointcloud/input_2048_4X/input_2048/
# up_rate = 4
# up_name = 4X

data_dir = ../local_distance_indicator/data/PU-GAN/test_pointcloud/input_2048_16X/input_2048/
up_rate = 16
up_name = 16X

To generate more uniform results for real scans (such as scanobjectnn), you can generate more uniform queries using the seed points in PU-SSAS and put them in "local_distance_indicator/data/ScanObjectNN/pussas_seed/". You need to set "use_seed_upsample = True" in the config, please refer to the testing of ScanObjectNN.

Evaluation

The evaluation is divided into two steps. We use the "PU-GAN 4X" setting for an example, and you need to modify the parameters for different settings.

First, use evaluation_code to compute P2F with:

python write_eval_script.py --dataset pugan --upsampled_pcd_path ../pretrained_model/pugan/test/4X/
bash eval_pugan.sh

Then, you need to shift to the virtual environment and code folder of PU-GCN, and then run the evaluate.py script of PU-GCN repo.

python evaluate.py \
--pred APU_LDI/global_field/pretrained_global/pugan/4X/ \ 
--gt APU_LDI/local_distance_indicator/data/PU-GAN/test_pointcloud/input_2048_4X/gt_8192 \
--save_path APU_LDI/global_field/pretrained_global/pugan/4X/

Train

To train APU-LDI on PU-GAN dataset, you can simply run the following command:

# training local ldi
cd local_distance_indicator
python train_ldi.py --dataset pugan

# learning global field
cd ../global_field
python train_global.py --conf confs/pugan.conf --mode train --dir pugan --dataset pugan --listname pugan.txt

And you can modify the config file to set training parameters.

Acknowledgments

Our code is built upon the following repositories: Grad-PU, PU-GCN, PU-GAN. Thanks for their great work.

Citation

If you find our code or paper useful, please consider citing

@inproceedings{li2024APU-LDI,
  title={Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling},
  author={Li, Shujuan and Zhou, Junsheng and Ma, Baorui and Liu, Yu-Shen and Han, Zhizhong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2024}
}

lisj575/APU-LDI