murphysong / pointcloudsegmentation

Experiments on point cloud segmentation.

Point Cloud Segmentation for Classific Feature Learning

Introduction

This work aims to show whether learning a point cloud segmentation task is able to extract features performing well in classification. We do all experiments under the framework of DGCNN.

Details for DGCNN see Dynamic Graph CNN for Learning on Point Clouds (https://arxiv.xilesou.top/pdf/1801.07829). We provide a PyTorch reimplementation for DGCNN in AnTao97/dgcnn.pytorch.

DGCNN provides two type of networks, one for classification and one for segmentation. We use "DGCNN_Cls" to denote network for classification and "DGCNN_Seg" for segmentation. The network sturcture of DGCNN is

We also do experiments to see whether learning segmentation on meaningful point clouds assembled by some base point clouds can help to learning better features for base point clouds. To run the experments, we first seperate segmentation parts of each shape from ShapeNetPart dataset as new shapes and make this new dataset as ShapeNetPart Part dataset. Then, we train network on ShapeNetPart dataset as normal and test it on ShapeNetPart Part dataset.

Some visualized point clouds in our ShapeNetPart Part dataset:

      chair           skateboard           airplane

Experimental results show that learning point cloud segmentation does help to extract features suitable for classification.

The key contributions of this work are as follows:

• Since the network provided by DGCNN for segmentation is supervised, we provide an revised DGCNN segmentation network with no category label.
• When segmentation is trained and tested on intact point clouds, the trained model can help to extract better features.
• When segmentation is trained on intact point clouds and tested on base point clouds, the trained model also can help to extract better features.

If you find this work useful, please cite:

@article{tao2020,
      Author = {An Tao},
      Title = {Point Cloud Segmentation for Classific Feature Learning},
      Journal = {https://github.com/AnTao97/PointCloudSegmentation},
      Year = {2020}
}

For more our works on point cloud feature learning, see this repo.

 

Requirements

• Python 3.7
• PyTorch 1.2
• CUDA 10.0
• Package: glob, h5py, tensorflow, tensorboard, tensorboardX and sklearn

 

Download datasets

Download the HDF5 format datasets (where each shape is sampled 2,048 points uniformly):

• ShapeNetPart (338M) [Dropbox] [BaiduDisk]
• ShapeNetPart Part (450M) [Dropbox] [BaiduDisk]
• ModelNet40 (194M) [Dropbox] [BaiduDisk]

You can find more details about the above datasets in this repo.

 

Experiment settings

To evaluate the quality of extracted features, we use ShapeNetPart dataset to both train DGCNN and a linear SVM classifier. Specifically, we train the linear SVM classifier on ShapeNetPart dataset using the features (latent representations) obtained from the trained feature encoder.

For transfer performance, we train the linear SVM classifier on ModelNet 40 dataset using the features (latent representations) obtained from the same network trained from the ShapeNetPart dataset.

In this work we compare the performance for adopted training task among supervised segmentation, supervised segmentation without category label, supervised classification and unsupervised reconstruction. For supervised segmentation without category label, we discard the adding of categorical vector (mlp {64}) and directly repeat the 1024 dim feature into n x 1024. We do unsupervised reconstruction following the framework in this repo and use source points from sphere surface for decoder. We also change feature dimension into 1024. Except unsupervised reconstruction, we do all experiments under the framework of DGCNN.

To train the network, run

python main.py --exp_name <exp name> --task <segment | classify | reconstruct> --dataset_root <root directory for datasets> --encoder <dgcnn_cls | dgcnn_seg> --k <20 | 40> --feat_dims 1024 --batch_size <16 | 32> --dataset shapenetpart --gpu <gpu ids>

Use --seg_no_label if you want to run segmentation task without category label.

You can download our already trained models from [Dropbox] or [BaiduDisk] and place them under snapshot/.

Because this work is done before our PyTorch reimplementation for DGCNN, the training setting in this repo is slightly different with AnTao97/dgcnn.pytorch. You can set self.epochs = 200 and random_translate=False in line 61 and 120 in segmentation.py to follow the settings in AnTao97/dgcnn.pytorch.

Besides ShapeNetPart dataset, we also test the performace of linear SVM classifier on ShapeNetPart Part dataset, using the model trained on ShapeNetPart dataset with segmentation task.

To evaluate the performance of a given trained model, run

python main.py --eval --model_path <model path> --task <segment | classify | reconstruct> --dataset_root <root directory for datasets> --encoder <dgcnn_cls | dgcnn_seg> --k <20 | 40> --feat_dims 1024 --dataset <shapenetpart | modelnet40 | shapenetpartpart> --gpu <gpu ids> 

Use --no_cuda if you want to run in CPU.

To use Tensorboard, run

tensorboard --logdir tensorboard --bind_all

You can find the Tensorboard records under tensorboard/.

 

Classification accuracy of linear SVM classifier

Results with best settings

Task	Info	Encoder	K	Batch Size	Epochs	ShapeNetPart	ModelNet40
Segmentation	Supervised	DGCNN_Seg	40	32	250	98.9%	89.4%
Segmentation	Supervised (no category label)	DGCNN_Seg	40	32	250	99.9%	89.2%
Classification	Supervised	DGCNN_Cls	40	32	250	99.8%	89.6%
Reconstruction	Unsupervised	DGCNN_Cls	20	16	250	98.7%	89.8%

 

 

Results with settings for segmentation task

Task	Info	Encoder	K	Batch Size	Epochs	ShapeNetPart	ModelNet40
Segmentation	Supervised	DGCNN_Seg	40	32	250	98.9%	89.4%
Segmentation	Supervised (no category label)	DGCNN_Seg	40	32	250	99.9%	89.2%
Classification	Supervised	DGCNN_Seg	40	32	250	99.9%	86.8%
Reconstruction	Unsupervised	DGCNN_Seg	40	32	290	98.8%	89.2%

 

 

Results evaluated on ShapeNetPart Part dataset

Task	Info	Encoder	K	Batch Size	Epochs	Training Dataset	Eval Acc
Segmentation	Supervised	DGCNN_Seg	40	32	250	ShapeNetPart	85.0%
Segmentation	Supervised (no category label)	DGCNN_Seg	40	32	250	ShapeNetPart	84.0%
Classification	Supervised	DGCNN_Cls	40	32	250	ShapeNetPart Part	99.0%
Reconstruction	Unsupervised	DGCNN_Cls	20	16	250	ShapeNetPart Part	87.5%

 

 

Performance analysis

Experimental results show that learning point cloud segmentation does help to extract features suitable for classification. However, simply adopting the training scheme from DGCNN for segmentation task is not suitable for transfer learning. We believe better results will get if using better training scheme.