YOLOv8-BeyondEarth

YOLOv8-BeyondEarth is a repository that contains tools and scripts to create YOLOv8 custom datasets, train the model, and post-process the obtained results. There is a strong focus on the use of satellite imagery and the application of the YOLOv8 model on solid planetary bodies in our Solar system.

To do

• Adapt get_sliced_prediction for use with Mask R-CNN.
• Use of batches for prediction could speed up the prediction.

Installation

Create a new environment if wanted. Then you can install the rastertools by writing the following in your terminal. Before installing YOLOv8-BeyondEarth, you need to install rastertools and shptools.

rastertools

git clone https://github.com/astroNils/rastertools.git
cd rastertools
python -m pip install --index-url https://test.pypi.org/simple/ --no-deps rastertools_BOULDERING
pip install -r requirements.txt

shptools

git clone https://github.com/astroNils/shptools.git
cd shptools
python -m pip install --index-url https://test.pypi.org/simple/ --no-deps shptools_BOULDERING
pip install -r requirements.txt

YOLOv8-BeyondEarth

git clone https://github.com/astroNils/YOLOv8-BeyondEarth
cd YOLOv8-BeyondEarth
python -m pip install --index-url https://test.pypi.org/simple/ --no-deps YOLOv8BeyondEarth
pip install -r requirements.txt

sahi

To get the last version of sahi.

git clone https://github.com/obss/sahi.git
cd sahi
pip install -e . 

You should now have access to this module in Python.

python

from YOLOv8BeyondEarth.predict import get_sliced_prediction

Bugs

If in your coordinate system, the latitude of origin is different than 0 degrees and this is expressed by the parameter ["Latitude of natural origin", there is a bug in the alignment of the prediction shapefiles (i.e., the shapefiles do not align with the raster/image). This bug does not happen for the Lunar Reconnaissance Orbiter NAC images. I wonder if this is due to the previously mentioned coordinate system/projection parameter and some bugs with rasterio/shapely? (need to do more digging there). When this parameter is not defined (i.e., a latitude of origin equal to 0), then there is no bug. Note that in NAC images, an equirectangular projection can be centered on a latitude without causing a bug. Looking at the coordinate system, the centering on the latitude seems to be described by another parameter (["Latitude of 1st standard parallel"). Therefore, one solution could be to convert the coordinate system so that the centered latitude is described by the 1st standard parallel (or just keep the center latitude at 0 degrees). You can switch between coordinate systems with the help of gdalwarp (https://gdal.org/programs/gdalwarp.html).

gdalwarp -s_src <old_coordinate_system> -t_srs <wkt_string_where_only_latitude_1st_standard_parallel_is_used> <src_raster> <dst_raster>

You can get information about the old coordinate system with the help of gdal_info.

gdalinfo <src_raster>

Getting Started

A jupyter notebook is provided as a tutorial (GET STARTED HERE).

Few links (YOLO, YOLOv8, anchor-free object detection)

https://github.com/ultralytics/ultralytics

https://medium.com/cord-tech/yolov8-for-object-detection-explained-practical-example-23920f77f66a

https://keylabs.ai/blog/comparing-yolov8-and-yolov7-whats-new/

https://learnopencv.com/fcos-anchor-free-object-detection-explained/

https://learnopencv.com/centernet-anchor-free-object-detection-explained/#What-is-Anchor-Free-Object-Detection

https://www.datacamp.com/blog/yolo-object-detection-explained