Skip to content

A Pytorch implementation of Generative Adversarial Network for Heuristics of Sampling-based Path Planning

License

Notifications You must be signed in to change notification settings

yuqikong/pathgan

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PathGAN

A Pytorch implementation of Generative Adversarial Network for Heuristics of Sampling-based Path Planning

Original arXiv paper

    .
    ├── assets                  # Images, results
    ├── data                    # Dataset location & data generation
    │   ├── movingai
    │   ├── map_augmentator.py  
    │   ├── task_generator.py   
    │   ├── roi_generator.py 
    │   ├── rrt.py
    │   ├── utils.py 
    │   └── dataset
    ├── gan                     # Scripts for GANs (original & pix2pix)
    │   ├── generator.py        
    │   ├── discriminator.py    
    │   ├── modules.py          
    │   ├── checkpoint            
    │   ├── utils
    |   |   ├── data.py
    |   |   ├── criterion.py
    |   |   ├── metrics.py
    |   |   └── trainer.py
    │   └── pix2pix
    |       ├── blocks.py
    |       ├── trainer.py
    |       └── utils.py
    ├── path                    # Scripts for pathfinding
    │   ├── rrt.py              
    │   ├── rrt_star.py         
    │   ├── RRT_updated.py 
    │   └── RRT_.py  
    ├── train.py                # To train GAN from the paper 
    ├── train_pix2pix.py        # To tran our Pix2Pix GAN 
    ├── roi_generation.ipynb    # To generate ROIs
    ├── LICENSE
    └── README.md

Table of content

Introduction

The overall structure of the PathGAN consists of two things:

  1. RRT* searching algorithm and
  2. Generative Aversarial Network for promising region generation

Searching algorithm

Path searching by RRT*

GAN architecture

Dataset

Dataset consists of 10,000 samples (Map, Point, ROI):

Alt text

  • train set (8,000 samples)
  • test set (2,000 samples)

Link to the dataset

Results

GAN Generated ROI's mIoU mDICE mFID mIS # of params * Checkpoint *
Original (from paper) Link 70.2% 82.0% 79.7 1.019 21,231,827 Link
Pix2Pix (ours) Link 58.1% 72.2% 91.2 1.017 4,170,477 Link

*- of Generator

  • mIoU - average Intersection over Union for all 2,000 samples in test set

  • mDICE -average DICE for all 2,000 samples in test set

  • mFID -average Frechet Inception Distance for all 2,000 samples in test set

  • mIS - average Inception Score for all 250 batches (2,000 samples/8 samples per batch) in test set

Original Generator (from paper)

Alt text

RRT result--> True ROI --> Pred ROI -->True ROI ∩ Pred ROI

Pix2Pix Generator (ours)

Alt text

RRT result--> True ROI --> Pred ROI -->True ROI ∩ Pred ROI

Results on MovingAI maps

GAN Generated ROI's mIoU mDICE mFID mIS
Original (from paper) Link 38.4% 53.8% 88.1 1.014
Pix2Pix (ours) Link 30.8% 46.3% 100.1 1.012

*- of Generator

  • mIoU - average Intersection over Union for all 699 samples

  • mDICE -average DICE for all 699 samples

  • mFID -average Frechet Inception Distance for all 699 samples

  • mIS - average Inception Score for all 88 batches (699 samples/8 samples per batch)

Original Generator (from paper)

Alt text

True ROI --> Pred ROI -->True ROI ∩ Pred ROI

Pix2Pix Generator (ours)

Alt text

True ROI --> Pred ROI -->True ROI ∩ Pred ROI

License

This project is licensed under MIT.

Links

About

A Pytorch implementation of Generative Adversarial Network for Heuristics of Sampling-based Path Planning

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • HTML 99.8%
  • Other 0.2%