• PCN: Point Completion Network.
• • Algorithm analysis
  • Running commands
  • Quantitative Result Comparison
  • • Seen Categories:
    • Unseen Categories
    • Seen Categories:
    • Unseen Categories
  • Contributors

PCN: Point Completion Network.

PCN is a point cloud completion algorithm based on deep learning, which is collected into our open source algorithm library of point cloud. The original project is run on python-3 and tensorflow-1.12 with CUDA 9.0 and tested on Ubuntu 16.04 with Python 3.5, which is in folder './PCN-master'. We add the PyTorch version which is in folder './PCN-PyTorch'.

Algorithm analysis

This project designes a deep learning architecture which combines advantages from existing architectures to generate a dense point cloud in a coarse-to-fine fashion, enabling high resolution completion with much fewer parameters than voxelbased models.

Running commands

Tensorflow

1. Install environment

pip install -r requirments.txt
pip install open3d

2. Train
   Download training (train.lmdb, train.lmdb-lock) and validation (valid.lmdb, valid.lmdb-lock) data from shapenet or shapenet_car directory on Google Drive

python3 train.py

3. Test

python3 test_shapenet.py

Note: you should revise codes in test_shapenet.py to choose test.list or test_novel.list

PyTorch

1. Built environment on Python-3.6, PyTorch-1.7, and CUDA10.1

2. Compile for cd and emd

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

3. Training

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

4. Test

python test.py --exp_name PCN_16384 --ckpt_path checkpoint/best_l1_cd.pth --batch_size 32 --num_workers 8

Quantitative Result Comparison

We compare the algorithm performance of different frameworks under tensorflow and PyTorch respectively. The main evaluation metric is CD. Note: we remove emd and F-score because of some bugs.

Tensorflow:

Seen Categories:

Category	L1_CD(1e-3)	L2_CD(1e-4)
Airplane	5.5025	1.4003
Cabinet	10.6254	4.4503
Car	8.6960	2.4445
Chair	10.9983	4.8382
Lamp	11.3389	6.2384
Sofa	11.6759	5.1293
Table	8.5901	3.5689
Vessel	9.6649	4.0620
Average	9.6365	4.0165

Unseen Categories

Category	L1_CD(1e-3)	L2_CD(1e-4)
Bus	9.4599	3.5176
Bed	21.6380	26.4793
Bookshelf	14.7976	12.1413
Bench	11.0292	5.9637
Guitar	10.4048	4.9493
Motorbike	14.7507	7.8966
Skateboard	12.0444	6.6281
Pistol	14.2382	9.3584
Average	13.5453	9.6168

PyTorch:

Seen Categories:

Category	L1_CD(1e-3)	L2_CD(1e-4)
Airplane	6.0028	1.7323
Cabinet	11.2092	4.7351
Car	9.1304	2.7157
Chair	12.0340	5.8717
Lamp	12.6754	7.5891
Sofa	12.8218	6.4572
Table	9.8840	4.5669
Vessel	10.1603	4.2766
Average	10.4897	4.7431

Unseen Categories

Category	L1_CD(1e-3)	L2_CD(1e-4)
Bus	10.5110	4.4648
Bed	24.9320	32.4809
Bookshelf	15.8186	13.1783
Bench	12.1345	7.3033
Guitar	11.4964	5.9601
Motorbike	15.3426	8.7723
Skateboard	13.1909	7.9711
Pistol	17.4897	15.5062
Average	15.1145	11.9546

Contributors

name: Zhang Yongchi
email: zhangych02@pcl.ac.cn