SO-Net: Self-Organizing Network for Point Cloud Analysis. CVPR 2018, Salt Lake City, USA Jiaxin Li, Ben M. Chen, Gim Hee Lee, National University of Singapore
SO-Net is a deep network architecture that processes 2D/3D point clouds. It enables various applications including but not limited to classification, shape retrieval, segmentation, reconstruction. The arXiv version of SO-Net can be found here.
@article{li2018sonet,
title={SO-Net: Self-Organizing Network for Point Cloud Analysis},
author={Li, Jiaxin and Chen, Ben M and Lee, Gim Hee},
journal={arXiv preprint arXiv:1803.04249},
year={2018}
}
Inspired by Self-Organizing Network (SOM), SO-Net performs dimensional reduction on point clouds and extracts features based on the SOM nodes, with theoretical guarantee of invariance to point order. SO-Net explicitly models the spatial distribution of points and provides precise control of the receptive field overlap.
This repository releases codes of 4 applications:
Requirements:
Optional dependency:
sudo pip3 install numba
4. Set environment variables, example:export LLVM_CONFIG=/usr/lib/llvm-6.0/bin/llvm-config
export NUMBAPRO_NVVM=/usr/local/cuda/nvvm/lib64/libnvvm.so
export NUMBAPRO_LIBDEVICE=/usr/local/cuda/nvvm/libdevice
For ModelNet40/10 and ShapeNetPart, we use the pre-processed dataset provided by PointNet++ of Charles R. Qi. For SHREC2016, we sampled points uniformly from the original *.obj
files. Matlab codes that perform sampling is provided in data/
.
In SO-Net, we can decouple the SOM training as data pre-processing. So we further process the datasets by generating a SOM for each point cloud. The codes for batch-SOM training can be found in data/
.
In addition, our prepared datasets can be found in Google Drive: MNIST, ModelNet, ShapeNetPart, SHREC2016.
The 4 applications share the same SO-Net architecture, which is implemented in models/
. Typically each task has its own folder like modelnet/
, part-seg/
that contains its own configuration options.py
, training script train.py
and testing script test.py
.
To run these tasks, you may need to set the dataset type and path in options.py
, by changing the default value of --dataset
, --dataroot
.
We use visdom for visualization. Various loss values and the reconstructed point clouds (in auto-encoder) are plotted in real-time. Please start the visdom server before training, otherwise there will be warnings/errors, though the warnings/errors won't affect the training process.
python3 -m visdom.server
The visualization results can be viewed in browser with the address of:
http://localhost:8097
Point cloud classification can be done on ModelNet40/10 and SHREC2016 dataset. Besides setting --dataset
and --dataroot
, --classes
should be set to the desired class number, i.e, 55 for SHREC2016, 40 for ModelNet40 and 10 for ModelNet10.
cd modelnet/
python3 train.py
The training of shape retrieval is the same as classification, while at testing phase, the score vector (length 55 for SHREC2016) is regarded as the feature vector. We calculate the L2 feature distance between each shape in the test set and all shapes in the same predicted category from the test set (including itself). The corresponding retrieval list is constructed by sorting these shapes according to the feature distances.
cd shrec16/
python3 train.py
Segmentation is formulated as a per-point classification problem.
cd part-seg/
python3 train.py
An input point cloud is compressed into a feature vector, based on which a point cloud is reconstructed to minimize the Chamfer loss. Supports ModelNet, ShapeNetPart, SHREC2016.
cd autoencoder/
python3 train.py
This repository is released under MIT License (see LICENSE file for details).
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。