Pytorch Implementation of PointNet and PointNet++

This repo is implementation for PointNet and PointNet++ in pytorch.

Data Preparation

• Download ModelNet here for classification and ShapeNet here for part segmentation. Uncompress the downloaded data in this directory. ./data/ModelNet and ./data/ShapeNet.
• Run download_data.sh and download prepared S3DIS dataset for sematic segmantation and save it in ./data/indoor3d_sem_seg_hdf5_data/

Classification

PointNet

• python train_clf.py --model_name pointnet

PointNet++

• python train_clf.py --model_name pointnet2

Performance

Model	Accuracy
PointNet (Official)	89.2
PointNet (Pytorch)	89.4
PointNet++ (Official)	91.9
PointNet++ (Pytorch)	91.8

• Training Pointnet with 0.001 learning rate in SGD, 24 batchsize and 141 epochs.
• Training Pointnet++ with 0.001 learning rate in SGD, 12 batchsize and 45 epochs.

Part Segmentation

PointNet

• python train_partseg.py --model_name pointnet

PointNet++

• python train_partseg.py --model_name pointnet2

Performance

Model	Inctance avg	Class avg	aero	bag	cap	car	chair	ear phone	guitar	knife	lamp	laptop	motor	mug	pistol	rocket	skate board	table
PointNet (Official)	83.7	80.4	83.4	78.7	82.5	74.9	89.6	73	91.5	85.9	80.8	95.3	65.2	93	81.2	57.9	72.8	80.6
PointNet (Pytorch)	82.4	78.4	81.1	77.8	83.7	74.3	83.3	65.7	90.5	85.1	78.1	94.5	63.7	91.7	80.5	56.2	73.7	67.5
PointNet++ (Official)	85.1	81.9	82.4	79	87.7	77.3	90.8	71.8	91	85.9	83.7	95.3	71.6	94.1	81.3	58.7	76.4	82.6
PointNet++ (Pytorch)	84.1	81.6	82.6	85.7	89.3	78.1	86.8	68.9	91.6	88.9	83.9	96.8	70.1	95.7	82.8	59.8	76.3	71.1

• Training both Pointnet and Pointnet++ with 0.001 learning rate in Adam, 16 batchsize, about 130 epochs and 0.5 learning rate decay every 20/30 epochs.
• Class avg is the mean IoU averaged across all object categories, and inctance avg is the mean IoU across all objects.
• In official version PointNet, author use 2048 point cloud in training and 3000 point cloud with norm in testing. In official version PointNet++, author use 2048 point cloud with its norm (Bx2048x6) in both training and testing.

Semantic Segmentation

PointNet

• python train_semseg.py --model_name pointnet

PointNet++

• python train_semseg.py --model_name pointnet2

Performance (test on Area_5)

Model	Mean IOU	ceiling	floor	wall	beam	column	window	door	chair	tabel	bookcase	sofa	board	clutter
PointNet (Official)	41.09	88.8	97.33	69.8	0.05	3.92	46.26	10.76	52.61	58.93	40.28	5.85	26.38	33.22
PointNet (Pytorch)	44.43	91.1	96.8	72.1	5.82	14.7	36.03	37.1	49.36	50.17	35.99	14.26	33.9	40.23
PointNet++ (Official)	N/A
PointNet++ (Pytorch)	52.28	91.7	95.9	74.6	0.1	18.9	43.3	31.1	73.1	65.8	51.1	27.5	43.8	53.8

• Training Pointnet with 0.001 learning rate in Adam, 24 batchsize and 84 epochs.
• Training Pointnet++ with 0.001 learning rate in Adam, 12 batchsize and 67 epochs.

Visualization

Using show3d_balls.py

cd visualizer
bash build.sh #build C++ code for visualization

Using pc_utils.py

TODO

• PointNet and PointNet++
• Experiment
• Visualization Tool

Reference By

halimacc/pointnet3
fxia22/pointnet.pytorch

Links

Official PointNet and Official PointNet++