This repository contains ongoing efforts for semantic segmentation for Robocup SPL.

• clone the BHuman dataset from https://sibylle.informatik.uni-bremen.de/public/datasets/semantic_segmentation/dataset.zip and unzip it into the repository folder
• create a python 3.10 virtual environment with python -m venv .venv
• enter the virtual environment (. .venv/bin/activate in bash)
• install all dependencies with pip install -r requirements.txt
• perform the train/test split with python -m src.split_test (this will perform a 90/10 train/test split and put the results into the datasplit folder)

You can now train the network with python -m src.main. All relevant global configuration can be found in src.configuration.constants. The network is initialized in the LightningWrapper and can be substituted there. The data augmentation pipeline in use is src.augmentation.augmenter2.

Hyperparameter search can be started with python -m src.hyperparameter_search {STUDY_NAME}, however I used the script start_hyperparameter_search.sh which starts two processes for hyperparameter search on two GPUs since I ran into issues doing this in one process.