Phase 1: Postprocess the semantic segmentations to obtain filtered instance segmentations

Phase 2: Evaluate instance seg against manual count for every image
You know the drill: setup a venv to isolate your environment (using conda, pyenv, ...). For example:
conda create python==3.12.9 -n eval-obj-det
conda activate eval-obj-det
conda env update --file environment.yaml
This should install AxonDeepSeg so we can use its utilities to postprocess the semantic segmentations.