Chainer Pointnet Save
Chainer implementation of PointNet, PointNet++, KD-Network and 3DContextNework
chainer-pointnet
Various point cloud based deep neural network implementation by Chainer [1].
It includes PointNet, PointNet++, Kd-Network and Kd context net (3DContextNet).
Installation
Please install Chainer (and cupy if you want to use GPU) beforehand,
# Install chainer
pip install chainer
# Install cupy if you use GPU (please change based on your CUDA version)
# For example, if you use CUDA 8.0,
pip install cupy-cuda80
# similarly cupy-cuda90 or cupy-cuda91 is avaialble.
and then type following to install master branch.
git clone https://github.com/corochann/chainer-pointnet.git
pip install -e chainer-pointnet
Also, some extension library is used in some of the code,
# Chainer Chemistry
git clone https://github.com/pfnet-research/chainer-chemistry.git
pip install -e chainer-chemistry
# ChainerEX
git clone https://github.com/corochann/chainerex.git
pip install -e chainerex
Model structure
PointNet [2]
Implementations are in models/pointnet.
Both classification and segmentation network are implemented.
models/pointnet_cls can be used for classification task.
trans option represents to use TransformNet or not.
trans=False corresponds PointNetVanilla (basic),
and trans=True corresponds PointNet in the paper, respectively.
In my experiment PointNetVanilla performs already very well,
the gain in PoinetNet is few (maybe only 1-2% gain) while computation becomes
much huge (around 3 times slower).
Original implementation (in tensorflow) can be found on github under MIT license.
tips
I found the batch normalizations in the last linear layers are quite important. The accuracy dramatically changes (10% or more) with BatchNormalization at FC layers.
PointNet++ [3]
Implementations are in models/pointnet2.
Both classification and segmentation network are implemented.
Just note that thanks to cupy,
there is no C language implementation for farthest point sampling, grouping or
feature propagation on GPU.
You don't need to build any C binary to use this function.
Please refer utils/sampling.py and utils/grouping.py for sampling & grouping.
Original implementation (in tensorflow) can be found on github under MIT license.
Kd-Network [4]
Implementations are in models/kdnet
Both classification and segmentation network are implemented.
Original implementation constructs KDTree by their own implementation, which supports random splitting.
However in my implementation, scipy.spatial.cKDTree is used for
easy and faster implementation, by following fxia22/kdnet.pytorch.
So implementation in this repo does not support random tree splitting.
Original implementation (in tensorflow) can be found on github under MIT license.
Also, pytorch implementations can be found on github
Kd Context Network (3DContextNet) [7]
Originally called 3DContextNet, but I named KDContextNet in my python program.
Implementations are in models/kdcontextnet
Both classification and segmentation network are implemented.
I could not find the implementation so far, and this is "inferred" implementation. Especially for segmentation part, how to "upsample" is not written in detail. So this implementation might be different from author's implementation.
Experiments
Experiments in each dataset is located under expriments folder.
Each folder is independent, so you can refer independently.
ModelNet40 [5]
This is point cloud classification task of 40 category. Download script & code is from charlesq34/pointnet
The dataset is automatically downloaded and preprocessed.
Dataset shape is train: (9840, 3, num_point=1024, 1), test (2468, 3, num_point=1024, 1).
ch=3, meaning it only contains (x, y, z) coordinate information.
You can simply execute train code to train PointNet or PointNetVanilla.
# use gpu with id 0, train PointNetVanilla
$ python train.py -g 0 --trans=false --method=point_cls --out=results/point_vanilla
# use gpu with id 0, train PointNet
$ python train.py -g 0 --method=point_cls --out=results/point
# use gpu with id 0, train PointNet++
$ python train.py -g 0 --method=point2_cls_ssg --out=results/point2_ssg
$ python train.py -g 0 --method=point2_cls_msg --out=results/point2_msg
# use gpu with id 0, train KDNet
$ python train.py -g 0 --method=kdnet_cls --dropout_ratio=0.3 --use_bn=1 --out=results/kdnet
# use gpu with id 0, train KDContextNet
$ python train.py -g 0 --method=kdcontextnet_cls --dropout_ratio=0.3 --use_bn=1 --out=results/kdcontextnet
python train.py -g 0 --method=kdcontextnet_cls --dropout_ratio=0.3 --use_bn=1 --out=results/kdcontextnet_level369
python train.py --use_bn=1 --dropout_ratio=0.3
# PointNetVanilla
epoch main/loss main/cls_loss main/trans_loss1 main/trans_loss2 main/accuracy validation/main/loss validation/main/accuracy lr elapsed_time
250 0.111644 0.111644 0.958367 0.606223 0.872596 1e-05 3560.81
# PointNet
250 0.119699 0.117399 0.00227531 2.40329e-05 0.95684 0.587751 0.871795 1e-05 10358
# PointNet2 SSG
250 0.0226786 0.989821 0.631495 0.898638 1e-05 54240.2
# PointNet2 MSG
250 0.0217696 0.991653 0.610621 0.892628 1e-05 160451
# KDNet with bn & dropout_ratio=0.3
250 0.10106 0.962235 1.01367 0.820913 1e-05 20324
# KDContextNet with bn & dropout_ratio=0.3
250 0.126861 0.952769 0.835642 0.825321 1e-05 31900.6
# KDContextNet with bn & dropout_ratio=0.3 & normalize=1 & residual=1
250 0.0804102 0.9716 0.824233 0.838542 1e-05 28028.5
KDNet seems "overfit" to the train data, meaning that its representation power is strong but it fails to generalize to test data. Maybe this is due to its property of non-rotational invariance.
S3DIS [6]
Stanford Large-Scale 3D Indoor Spaces Dataset (S3DIS) for point cloud segmentation task. Download the dataset from,
Prerocessing code adopted from charlesq34/pointnet
under third_party directory.
Dataset shape:
train shape (20291, 9, 4096, 1) (20291, 4096)
test shape (3294, 9, 4096, 1) (3294, 4096)
ch=9, it contains 0-2 ch: XYZ, 3-5ch: RGB, 6-8 normalized XYZ coordinate
information respectively.
Steps:
-
Go to download link: download S3DIS dataset from S3DIS Dataset. You need to send application form.
-
Download
Stanford3dDataset_v1.2_Aligned_Version.zipfile (4GB), place it unders3dis/datadirectory.
2'. Fix mis label manually.
Stanford3dDataset_v1.2_Aligned_Version/Area_5/hallway_6/Annotations/ceiling_1.txt
has wrong charcter at line 180389. Please fix it manually.
- Preprocessing
collect_indoor3d_data.py is for data re-organization and
gen_indoor3d_h5.py is to generate HDF5 files. (cite from charlesq34/pointnet)
$ cd third_party
$ python collect_indoor3d_data.py
$ python gen_indoor3d_h5.py
- Training
# use gpu with id 0, train PointNetVanilla
$ python train.py -g 0 --method=point_seg --trans=false --out=results/pointnet_vanilla
# use gpu with id 0, train PointNet
$ python train.py -g 0 --method=point_seg --out=results/pointnet
# use gpu with id 0, train PointNet++
$ python train.py -g 0 --method=point2_seg_ssg --out=results/pointnet2
# use gpu with id 0, train KDNet
$ python train.py -g 0 --method=kdnet_seg --dropout_ratio=0 --use_bn=1
$ python train.py -g 0 --method=kdnet_seg --dropout_ratio=0.3 --use_bn=1 --out=results/kdnet
# use gpu with id 0, train KDContextNet
$ python train.py -g 0 --method=kdcontextnet_seg --dropout_ratio=0.3 --use_bn=1 --out=results/kdcontextnet
python train.py --use_bn=1 --dropout_ratio=0.3
# PointNetVanilla
epoch main/loss main/cls_loss main/trans_loss1 main/trans_loss2 main/accuracy validation/main/loss validation/main/accuracy lr elapsed_time
250 0.0305516 0.0305516 0.988418 0.690604 0.891049 1e-05 101405
ScanNet
Point cloud semantic segmentation task of indoor scenes.
LICENSE
MIT License.
No warranty or support for this implementation. Each model performance is not guaranteed, and may not achieve the score reported in each paper. Use it at your own risk.
Please see the LICENSE file for details.
I appreciate the authors who open sourced their code for the reference under permissive license.
Reference
[1] Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in Advances in Neural Information Processing System (NIPS) 28, 2015.
[2] Qi, Charles R and Su, Hao and Mo, Kaichun and Guibas, Leonidas J. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. arXiv preprint arXiv:1612.00593, 2016.
[3] Qi, Charles R and Yi, Li and Su, Hao and Guibas, Leonidas J. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. arXiv preprint arXiv:1706.02413 2017.
[4] Roman, Klokov and Victor, Lempitsky. Escape from Cells: Deep Kd-Networks for the Recognition of 3D Point Cloud Models. arXiv preprint arXiv:1704.01222 2017.
[5] Z, Wu and S, Song and A, Khosla and F, Yu and L, Zhang and X, Tang and J, Xiao. 3D ShapeNets: A Deep Representation for Volumetric Shapes. Proceedings of 28th IEEE Conference on Computer Vision and Pattern Recognition (CVPR2015), 2015.
[6] Iro, Armeni and Alexander, Sax and Amir, R., Zamir and Silvio, Savarese. Joint 2D-3D-Semantic Data for Indoor Scene Understanding. arXiv preprint arXiv:1702.01105 2017.
[7] Wei, Zeng and Theo, Gevers. 3DContextNet: K-d Tree Guided Hierarchical Learning of Point Clouds Using Local and Global Contextual Cues. arXiv preprint arXiv:1711.11379 2017.
Open Source Agenda Badge
