This is the official implementation of the codes and data that produced the results in titled
"A Kernel-based Multi-featured Rock Modeling and Detection Framework for a Mars Rover" submitted to IEEE TNNLS Special Issue on: Deep Learning for Earth and Planetary Geosciences.
We will soon released the whole codes and data after the paper is availabe.
We refined and built a labeled dataset called MarsData for rock segmentation on planet excepically on Mars in this paper. Images are collected from an unlabeled Mars image dataset. The intention of MarsData is to provied a standard rock detection benchmark with pixel-level mask for researchers who are interseting on deep learning methods for planetray sciences and robotics. Note: All mars images are courtesy of NASA/JPL-Caltech. You can read the full use policy here.
Currently, MarsData currently includes two sub-datasets, Rock-A, Rock-B with total 405 labeled rock images and more than 20,000 rocks. Rock-A is a simple rock dataset with a few rocks in one scene. Rock-B is a challenging dataset with more abundant rocks in one image. We used them to evaluate proposed algorithms and others in our paper. In order to produce sufficient data to well support the deep training, we combined them together and split all images randomly into train and test sets. After data augmentation, the train and test sets can be used to train and evaluate the deep learning-based rock segmentation methods, as mentioned in our paper. Of course, you can do the augmentation work by yourself with more methods in order to produce more data.
| MarsData | Rock-A | Rock-B | train(after aug) | test(after aug) |
|---|---|---|---|---|
| Number of images | 201 | 204 | 1981 | 854 |
| Average rock number | 3.15 | 15.89 | - | 9.91 |
| max rock number | 12 | 77 | - | 55 |
| min rock number | 1 | 1 | 0 | 0 |
- Matlab R2018b and C++
-
Exfeatures
This folder includes all the functions that extract gray-scale features for each superpixel region. -
Dependencies
This folder includes the core functions of the proposed KPCA-based Rock Detection method(KPRD), KLRR-based Rock Detection method(KLRD) and the RKLRR method. In addition, it also includes codes for superpixel generation in folder of SLIC, as well as some dependent functions used in our algorithms. -
Evaluation
This folder includes all the source codes of the evaluating metrics used in our paper.
- clone the respository to your PC
- for testing KLRD, KPRD and RKLRR, please run demo.m, all the detection results will be saved in the folder of Results
- for evaluating, run evaluate.m
We will continue to improve MarsData, planing to build a more complex and challenging sub-dataset(tentatively named Rock-C), and striving to reach more than 2000 labeled images. Anyone who wants join us to further optimize and improve MardData is strongly welcome. Besides, we are applying for the data of China's First Mars Exploration Mission("TianWen-1"), looking for passionate researchers to collaborate with us! Contact: alexcapshow@gmail.com or meibaoyao@jlu.edu.cn.
If you find MarsData is helpful for your research, please cite our paper：
@article{xiao2021kernel,
title={A Kernel-Based Multi-Featured Rock Modeling and Detection Framework for a Mars Rover},
author={Xiao, Xueming and Yao, Meibao and Liu, Haiqiang and Wang, Jiake and Zhang, Lei and Fu, Yuegang},
journal={IEEE Transactions on Neural Networks and Learning Systems},
year={2021},
doi={10.1109/TNNLS.2021.3131206},
publisher={IEEE}
}