zhulf0804/pointnet2.pytorch

A PyTorch Implementation of Pointnet++.

Introduction

An unofficial PyTorch Implementation of PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space[NIPS 2017].

Requirements

• PyTorch, Python3, TensorboardX, tqdm, fire

Classification

• Start

  • Dataset: ModelNet40, download it from Official Site or Baidu Disk with hi1i.

  • Train

    python train_clss.py --data_root your_data_root --log_dir your_log_dir
    
    eg.
    python train_clss.py --data_root /root/modelnet40_normal_resampled --log_dir cls_ssg_1024
    

  • Evaluate

    python evaluate.py evaluate_cls model data_root checkpoint npoints
    
    eg.
    python evaluate.py evaluate_cls pointnet2_cls_ssg  /root/modelnet40_normal_resampled \
    checkpoints/pointnet2_cls_250.pth 1024
    
    python evaluate.py evaluate_cls pointnet2_cls_msg root/modelnet40_normal_resampled \
    checkpoints/pointnet2_cls_250.pth 1024
    

• Performance(the first row is the results reported in Paper, the following rows are results reported from this repo.)

  Model	NPoints	Aug	Accuracy(%)
  PointNet2(official)	5000	✓	91.7
  PointNet2_SSG	1024	✗	91.8
  PointNet2_SSG	4096	✗	91.7
  PointNet2_SSG	4096	✓	90.5
  PointNet2_MSG	4096	✓	91.0

  Model	Train_NPoints	DP	Test_NPoints	Accuracy(%)
  PointNet2_SSG	1024	✗	256	67.9
  PointNet2_SSG	1024	✓	256	90.8
  PointNet2_SSG	1024	✗	1024	91.8
  PointNet2_SSG	1024	✓	1204	91.9

• Train Your own Dataset

  • Prepare the dataset(n classes) in the ModelNet40 structure

    CustomData(dir)
        |- class1(dir)
            | - class1_name11.txt
            | - class1_name12.txt
            ...
        |- class2(dir)
            | - class2_name21.txt
            | - class2_name22.txt
            ...
        |- classn(dir)
        |- shape_names.txt
            | - class1(line1)
            | - class2(line2)
            | - ...
            | - classn(linen)
        |- train.txt
            | - class1_name11
            | - class2_name21
            | - class2_name22
            | - ...
            | - classn_namen1
        |- test.txt
            | - class1_name12
            | - class2_name22
            | - ...
            | - classn_namen2
    

  • Start to train

    python train_custom_cls.py --data_root your_datapath/CustomData --nclasses 2 --npoints 2048
    

  • Start to evaluate

    python evaluate_custom.py evaluate_cls pointnet2_cls_ssg your_datapath/CustomData work_dirs/checkpoints/pointnet2_cls_250.pth 2
    

Part Segmentation

• Start

  • Dataset: ShapeNet part, download it from Official Site or Baidu Disk with 3e5z.

  • Train

    python train_part_seg.py --data_root your_data_root --log_dir your_log_dir
    
    eg.
    python train_part_seg.py --data_root /root/shapenetcore_partanno_segmentation_benchmark_v0_normal \
    --log_dir seg_ssg --batch_size 64
    

  • Evaluate

    python evaluate.py evaluate_seg data_root checkpoint
    
    eg.
    python evaluate.py evaluate_seg /root/shapenetcore_partanno_segmentation_benchmark_v0_normal \
    seg_ssg/checkpoints/pointnet2_cls_250.pth
    

• Metrics: Average IoU

  Model	Metrics	mean	aero	bag	cap	car	chair	ear phone	guitar	knife	lamp	laptop	motor	mug	pistol	rocket	skate board	table
  PointNet2(official)	IoU	85.1	82.4	79.0	87.7	77.3	90.8	71.8	91.0	85.9	83.7	95.3	71.6	94.1	81.3	58.7	76.4	82.6
  PointNet2_SSG	IoU	84.1	82.3	75.0	80.1	77.8	90.2	73.7	90.7	84.1	82.9	95.0	69.3	93.3	80.3	55.6	76.3	80.7
  PointNet2_SSG	Accuracy	93.2	89.9	89.0	85.5	91.8	94.4	93.5	96.1	91.1	89.2	96.9	87.4	96.4	93.7	77.2	95.9	94.8

Reference

• https://github.com/charlesq34/pointnet2
• https://github.com/yanx27/Pointnet_Pointnet2_pytorch
• https://github.com/sshaoshuai/Pointnet2.PyTorch

pointnet2.pytorch's Issues

Formatting custom dataset

what would be the best to way to run newly acquired pointclouds through a trained model without needing to create a new custom dataset with the clouds each time? seems like it needs the train.txt shape_names.txt and test.txt redone on each new run

About normal

What should i do if i don‘t want to use normal?
Thanks for your answer.

Using code for just xyz attributes

Hi!
Many thanks for sharing your code. Could you please let me know how I can run your repository for point clouds that have just xyz attributes?
Thanks in advance.
Best regards,
Narges