SonarInterestPointDetection

Installation

Requirements

• Python 3 >= 3.5
• PyTorch >= 1.1
• OpenCV >= 3.4
• NumPy >= 1.18

conda create --name sonar-ip python=3.6
conda activate sonar-ip
pip install -r requirements.txt

Repository Structure

.
├── data                   
|   ├──demo-synthethic-shapes
|   └──sonar-data
|   |   └──test
|   |   |   ├──img
|   |   |   └──pts
|   |   └──train
|   |   |   ├──img
|   |   |   └──pts
|   |   └──val
|   |   |   ├──img
|   |   |   └──pts
|   └──annotation_to_npy.py
├── demo                   
|   ├──pretrained-superpoint.pth
|   └──sampled_trainer_notebook.ipynb
├── src                   
|   ├──data
|   |   ├──dataloader.py
|   |   └──dataset.py
|   ├──loss.py
|   ├──superpoint.py
|   ├──trainer.py
|   └──utils.py             
├── README.md
└── requirements.txt

How to run the file

.py files are available, however we advise you to run the demo notebook (.\demo\sample_trainer_notebook.py) in your Google Colab instance.

About the data

Data folder structure should follow above structure in order to run the training instance without running into any error. Data should be split into test, train, and validation folders each comprising img and pts folders. img should contain original .png files of your sonar image data, whereas pts should contain .npy files of your annotated interest points from the original image. We have used MakeSense.ai in order to manual annotate the interest points of the image files. Once you export .csv of your annotations, you can refer to .\data\annotation_to_npy.py file for the .npy conversion.

Example of img\*.png file

Example of pts\*.npy file

array ([[201, 404],
       [243, 439],
       [317, 415]])