https://github.com/garyli1019/pointnet-keras

Keras implementation for Pointnet
https://github.com/garyli1019/pointnet-keras

classification deep-learning keras point-cloud pointnet segmentation

Last synced: 25 days ago
JSON representation

Keras implementation for Pointnet

• Host: GitHub
• URL: https://github.com/garyli1019/pointnet-keras
• Owner: garyli1019
• License: mit
• Created: 2017-12-31T05:24:42.000Z (almost 7 years ago)
• Default Branch: master
• Last Pushed: 2019-10-28T03:41:10.000Z (about 5 years ago)
• Last Synced: 2024-01-21T10:22:38.176Z (11 months ago)
• Topics: classification, deep-learning, keras, point-cloud, pointnet, segmentation
• Language: Python
• Homepage:
• Size: 1.65 MB
• Stars: 122
• Watchers: 4
• Forks: 42
• Open Issues: 7
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# pointnet-keras

Original tensorflow implementation: https://github.com/charlesq34/pointnet

Package requirement: Python3.6, keras, tensorflow, numpy, matplotlib, h5py

## Results

Segmentation Sample

![seg_sample](images/seg_sample.png)

## How to Run code:

### Classification:

- Download the aligned dataset from [Link](https://shapenet.cs.stanford.edu/media/modelnet40_ply_hdf5_2048.zip)

- Put all traning `.h5` files under `Prepdata` folder, all testing `.h5` files under `Prepdata_test` folder

- Run train_cls.py. Accuracy rate will be 82.5%, which is slightly lower than the original implementation. 

### Segmentation:

- Download and unzip the shapenet dataset from [Link](https://shapenet.cs.stanford.edu/ericyi/shapenetcore_partanno_v0.zip).

- Run Seg_dataprep.py then train_seg.py.

## Point Architecture

- Input Transformation Net: Input: Nx3 point cloud sample, Output: 3x3 transformation net

![input_transformation_net](images/input_transformation_net.png)

- Feature Transformation Net:

![feature_transformation_net](images/feature_transformation_net.png)

- Global Feature: Input: Nx3 point cloud sample multiply input T_net. Output: 1*1024 global feature

- Classification Net: Input: Nx3 point cloud sample multiply input T_net. Output: 1x40 softmax prediction

![classification_net](images/classification_net.png)