Home

Awesome

<center>

SADMA: SAtellite baseD MArine debris detection

PWC <a href="#"><img src="https://img.shields.io/github/workflow/status/milesial/PyTorch-UNet/Publish%20Docker%20image?logo=github" /></a> <a href="https://pytorch.org/"><img src="https://img.shields.io/badge/PyTorch-v1.9.0-red.svg?logo=PyTorch" /></a> <a href="#"><img src="https://img.shields.io/badge/python-v3.6+-blue.svg?logo=python" /></a>

</center>

Implementation of the Residual Attention UNet in PyTorch on MARIDA. Marine Debris Archive (MARIDA) is a marine debris-oriented dataset on Sentinel-2 satellite images. It also includes various sea features that co-exist. MARIDA is primarily focused on the weakly supervised pixel-level semantic segmentation task. This repository hosts the basic tools for the extraction of spectral signatures as well as the code for the reproduction of the Residual Attention UNet model.

MARIDA can be easily downloaded from here. Alternatively, MARIDA can be downloaded from the Radiant MLHub.

<h2> Contents </h2>

Pre-print Paper

Link to arxiv pre-print: ResAttUNet: Detecting Marine Debris using an Attention activated Residual UNet If you find the repository useful, please consider giving a star ⭐ and cite:

@inproceedings{Mohammed2022ResAttUNetDM,
  title={ResAttUNet: Detecting Marine Debris using an Attention activated Residual UNet},
  author={Azhan Mohammed},
  year={2022}
}

Installing requirements

The requirements for running the code can be easily installed using the requirements.txt file. To install the requirements just use:

pip install -r requirements.txt

Once the requirements are installed, you can use the training and evaluation code easily.

Model details

The introduced model is a mixture of existing Residual blocks and Convolutional Block Attention Module with a simple UNet model. The residual blocks are used after the downsampling layers to extract deeper features, and channel and spatial attention is used to apply extra attention to the sparse nature of the debris. The blocks used in the network are shown in Figure 1

<div> <center>

Figure 1: Building blocks of ResAttUNet

Building blocks of ResAttUNet

</center> </div>

The complete network architecture is shown in Figure 2

<div> <center>

Figure 2: ResAttUNet Architecture

ResAttUNet Architecture

</center> </div>

Results

The introduced model achieves state-of-the-art results across all metrics on the test set of MARIDA. Figure 3 shows the performance of various models discussed, along with existing state-of-the-art results on the dataset.

<div> <center>

Figure 3: Results on test dataset

Results on test dataset

</center> </div>

Pre-trained Weights

The pre-trained weights for the model can be downloaded from Google drive

Author Details