SkySegmenatation

Report is here : https://clementstrauss.github.io/SkySeg/

Repository Structure :

00_Dataset

• Script to Download and convert coco panoptic to binary masks for the sky label. Only The Coco validation set is used as it is smaller in size.
• Notebook to load and display image from ADE20K, convert it to binary image masks

01_DeepLearning

• checkout the train and evaluation notebook, and the report
• should run fine with conda :

Edit 02/21/2023 : something is broken in the last fastai version (2.7.11)

conda create --name fast python=3.9
conda activate fast
conda install -c fastchan fastai==2.7.10
conda install jupyterlab
pip install opencv-python

02_AutoGrabCut_Cpp

• Computer vision based sky segmentation, run near real time on 720p video
• Developped in C++. It needs to be compiled. Instructions included in folder

TestImages, TestVideo

• images and video for subjective testing.
• Results on images are located in both 01_DeepLearning and 02_AutoGrabCut_Cpp
• Results on the video is located in 02_AutoGrabCut_Cpp

report.ipynb

• this is the notebook used to generate the report's web page, using quarto tool : https://quarto.org/