SaraR-1 / Aerial-YOLO-SAM

AerialML

This project aims to detect and segment buildings (and vehicles/aircraft as desired) from aerial imagery.

Specifically, the library is pipeline composed of the following steps:

Training:

• Load a training set of data - xView.
• Tile the training set images, increasing the pixels/meter downstream.
• Train a YOLOv8 model on this dataset to detect major types of objects.

Inference:

• Conduct image preprocessing to optimize input data quality (enhance contrast).
• Detect major types of objects such as buildings, cars or airplanes on our test dataset.
• Employ Segment Anything Model as a "prompt-based" segmentation technique.
• Generate a GeoJSON file containing polygon representations of the detected the object footprints.
• Create new images with the object footprints superimposed for visualization.

An illustration of the pipeline:

flowchart TD
        node1[("Input Data")]
        node2[("xView Data")]
        node4["Data Preparation"]
        node5["Segmentation"]
        node6["Object Detection"]
        node7["Object Detection Training"]
        node8["Pre-processing"]
        node9["Image Tiling"]
        node1-->node6
        node1-->node8
        node2-->node4
        node2-->node9
        node4-->node9
        node6-->node5
        node7-->node6
        node8-->node5
        node8-->node6
        node9-->node7

Installation

To use the code in this project, you'll need to create a Python environment, clone the repository and download the data using DVC.

conda create -n <envname> python=3.10
conda activate <envname>

pip install poetry
poetry install

git clone https://github.com/SaraR-1/AerialMI.git
dvc pull

Before using DVC, you'll first need to setup your AWS credential. To do so, create a file at ~/.aws/credentials. The content of the file should look like this:

[default]
aws_access_key_id=foo
aws_secret_access_key=bar

Where:

• aws_access_key_id - The access key for your AWS account.
• aws_secret_access_key - The secret key for your AWS account.

Please send me an email for these credentials.

Running the Code

To run the entire pipeline you can use DVC as:

dvc repro

This will reproduce the complete (or partial) pipeline by running its stages as needed in the correct order.