PCN: Point Completion Network

Introduction

This is implementation of PCN——Point Completion Network in pytorch. PCN is an autoencoder for point cloud completion. As for the details of the paper, please refer to arXiv.

Environment

• Ubuntu 18.04 LTS
• Python 3.7.9
• PyTorch 1.7.0
• CUDA 10.1.243

Prerequisite

Compile for cd and emd:

cd extensions/chamfer_distance
python setup.py install
cd ../earth_movers_distance
python setup.py install

Hint: Don't compile on Windows platform.

As for other modules, please install by:

pip install -r requirements.txt

Dataset

Please reference render and sample to create your own dataset. Also, we decompressed all .lmdb data from PCN data into .ply data which has smaller volume 8.1G and upload it into Google Drive. Here is the shared link: Google Drive.

Training

In order to train the model, please use script:

python train.py --exp_name PCN_16384 --lr 0.0001 --epochs 400 --batch_size 32 --coarse_loss cd --num_workers 8

If you want to use emd to calculate the distances between coarse point clouds, please use script:

python train.py --exp_name PCN_16384 --lr 0.0001 --epochs 400 --batch_size 32 --coarse_loss emd --num_workers 8

Testing

In order to test the model, please use follow script:

python test.py --exp_name PCN_16384 --ckpt_path <path of pretrained model> --batch_size 32 --num_workers 8

Because of the computation cost for calculating emd for 16384 points, I split out the emd's evaluation. The parameter --emd is used for testing emd. The parameter --novel is for novel testing data contains unseen categories while training. The parameter --save is used for saving the prediction into .ply file and visualize the result into .png image.

Pretrained Model

The pretrained model is in checkpoint/.

Results

I trained the model on Nvidia GPU 1080Ti with L1 Chamfer Distance for 400 epochs with initial learning rate 0.0001 and decay by 0.7 every 50 epochs. The batch size is 32. Best model is the minimum L1 cd one in validation data.

Quantitative Result

The threshold for F-Score is 0.01.

Seen Categories:

Category	L1_CD(1e-3)	L2_CD(1e-4)	EMD(1e-3)	F-Score(%)
Airplane	6.0028	1.7323	10.5922	86.2954
Cabinet	11.2092	4.7351	27.1505	61.6697
Car	9.1304	2.7157	14.3661	70.5874
Chair	12.0340	5.8717	22.4904	58.2958
Lamp	12.6754	7.5891	58.7799	57.8894
Sofa	12.8218	6.4572	19.2891	53.4009
Table	9.8840	4.5669	23.7691	70.9750
Vessel	10.1603	4.2766	17.9761	66.6521
Average	10.4897	4.7431	24.3017	65.7207

Unseen Categories

Category	L1_CD(1e-3)	L2_CD(1e-4)	EMD(1e-3)	F-Score(%)
Bus	10.5110	4.4648	17.0274	66.9774
Bed	24.9320	32.4809	42.7974	32.2265
Bookshelf	15.8186	13.1783	28.5608	50.0337
Bench	12.1345	7.3033	12.7497	62.4376
Guitar	11.4964	5.9601	28.4223	59.4976
Motorbike	15.3426	8.7723	21.8634	44.7431
Skateboard	13.1909	7.9711	17.9910	58.4427
Pistol	17.4897	15.5062	33.8937	45.6073
Average	15.1145	11.9546	25.4132	52.4958

Qualitative Result

Seen Categories

Unseen Categories

Citation

• PCN: Point Completion Network
• PCN's official Tensorflow implementation