ASFM-Net-Pytorch

Status: Done

This is an unofficial implement of ASFM-Net, and I have achieved CD-P 11.87 at PCN datasets.🚀

Official repo: https://github.com/Yan-Xia/ASFM-Net

Preparation

pytorch > 1.10.1 required!

1. Complie Chamfer3D, code from https://github.com/AllenXiangX/SnowflakeNet

cd Chamfer3D
python setup.py --user install

2. Complie pointnet++ which includes fps_sampling we need, code from https://github.com/AllenXiangX/SnowflakeNet

cd pointnet2_ops_lib
python setup.py --user install

3. Dataset path configs at config_pcn.py, modify it refer to https://github.com/hzxie/GRNet

__C.DATASETS.SHAPENET.PARTIAL_POINTS_PATH = '/home/chriskafka/dataset/ShapeNetCompletion/%s/partial/%s/%s/%02d.pcd'
__C.DATASETS.SHAPENET.COMPLETE_POINTS_PATH = '/home/chriskafka/dataset/ShapeNetCompletion/%s/complete/%s/%s.pcd'

Note that I am using gradient accumulation to simulate a big batchsize(32 in paper), You can set accumulation_steps = 1 in train_*.py under core directoroies to cancel this.

Run

0. Install and run tensorboard(optional):

https://www.tensorflow.org/tensorboard/get_started?hl=zh-cn

1. Step1: Pretrine PCN. bs 32, 206 epochs, CD Loss descended to 9.917

python main_pcn.py --baseline

2. Step2: Modify __C.CONST.PCNWEIGHTS in config_pcn.py to direct to pre-trained models obtained in step1,

__C.CONST.PCNWEIGHTS = 'checkpoint/pcn-baseline-best.pth'

train ASFM-net.

python main_pcn.py --backbone

bs 32, 116 epochs, CD Loss descended to 11.87(12.09 in paper).

Results

1. Coarse cd loss represention of step2, red for train loss while green for test.

2. Fine cd loss represention, green for test in step2 while blue for test in step1.

3. Feature Matching loss represention of step2.

4. Total loss represention of step2.

Thanks To

PCN pytorch implementation refers to https://github.com/paul007pl/VRCNet.

Point Clouds Utils(FPS, CD) borrowed from https://github.com/AllenXiangX/SnowflakeNet.

Data precessing codes comes from https://github.com/hzxie/GRNet.

And thanks ASFM-Net for their contributions.

If you have any suggestions or ideas about my work, feel free to refer a issue even a pr to keep me informed, thanks again.