DGCNN.pytorch

[中文版]

This repo is a PyTorch implementation for Dynamic Graph CNN for Learning on Point Clouds (DGCNN)(https://arxiv.xilesou.top/pdf/1801.07829). Our code skeleton is borrowed from WangYueFt/dgcnn.

Note that the network structure (Fig. 3) for classification in DGCNN paper is not consistent with the corresponding description in section 4.1 of the paper. The author of DGCNN adopts the setting of classification network in section 4.1, not Fig. 3. We fixed this mistake in Fig. 3 using PS and present the revised figure below.

Tip: The result of point cloud experiment usually faces greater randomness than 2D image. We suggest you run your experiment more than one time and select the best result.

Requirements

• Python 3.7
• PyTorch 1.2
• CUDA 10.0
• Package: glob, h5py, sklearn

Contents

• Point Cloud Classification
• Point Cloud Part Segmentation
• Point Cloud Semantic Segmentation

Point Cloud Classification

Run the training script:

• 1024 points

python main_cls.py --exp_name=cls_1024 --num_points=1024 --k=20 

• 2048 points

python main_cls.py --exp_name=cls_2048 --num_points=2048 --k=40 

Run the evaluation script after training finished:

• 1024 points

python main_cls.py --exp_name=cls_1024_eval --num_points=1024 --k=20 --eval=True --model_path=checkpoints/cls_1024/models/model.t7

• 2048 points

python main_cls.py --exp_name=cls_2048_eval --num_points=2048 --k=40 --eval=True --model_path=checkpoints/cls_2048/models/model.t7

Run the evaluation script with pretrained models:

• 1024 points

python main_cls.py --exp_name=cls_1024_eval --num_points=1024 --k=20 --eval=True --model_path=pretrained/model.cls.1024.t7

• 2048 points

python main_cls.py --exp_name=cls_2048_eval --num_points=2048 --k=40 --eval=True --model_path=pretrained/model.cls.2048.t7

Performance:

ModelNet40 dataset

Point Cloud Part Segmentation

Run the training script:

• Full dataset

python main_partseg.py --exp_name=partseg 

• With class choice, for example airplane

python main_partseg.py --exp_name=partseg_airplane --class_choice=airplane

Run the evaluation script after training finished:

• Full dataset

python main_partseg.py --exp_name=partseg_eval --eval=True --model_path=checkpoints/partseg/models/model.t7

• With class choice, for example airplane

python main_partseg.py --exp_name=partseg_airplane_eval --class_choice=airplane --eval=True --model_path=checkpoints/partseg_airplane/models/model.t7

Run the evaluation script with pretrained models:

• Full dataset

python main_partseg.py --exp_name=partseg_eval --eval=True --model_path=pretrained/model.partseg.t7

• With class choice, for example airplane

python main_partseg.py --exp_name=partseg_airplane_eval --class_choice=airplane --eval=True --model_path=pretrained/model.partseg.airplane.t7

Performance:

ShapeNet part dataset

Point Cloud Semantic Segmentation

The network structure for this task is slightly different with part segmentation, without spatial transform and categorical vector. The MLP in the end is changed into (512, 256, 13) and only one dropout is used after 256.

You have to download Stanford3dDataset_v1.2_Aligned_Version.zip manually from https://goo.gl/forms/4SoGp4KtH1jfRqEj2 and place it under data/

Run the training script:

This task use 6-fold training, such that 6 models are trained leaving 1 of 6 areas as the testing area for each model.

• Train in area 1-5

python main_semseg.py --exp_name=semseg_6 --test_area=6 

Run the evaluation script after training finished:

• Evaluate in area 6 after model is trained in area 1-5

python main_semseg.py --exp_name=semseg_eval_6 --test_area=6 --eval=True --model_root=checkpoints/semseg/models/

• Evaluate in all areas after 6 models are trained

python main_semseg.py --exp_name=semseg_eval --test_area=all --eval=True --model_root=checkpoints/semseg/models/

Run the evaluation script with pretrained models:

• Evaluate in area 6

python main_semseg.py --exp_name=semseg_eval_6 --test_area=6 --eval=True --model_root=pretrained/semseg/

• Evaluate in all areas

python main_semseg.py --exp_name=semseg_eval --test_area=all --eval=True --model_root=pretrained/semseg/

Performance:

Stanford Large-Scale 3D Indoor Spaces Dataset (S3DIS) dataset