Skip to content
/ hapt3D Public

3D Hierarchical Panoptic Segmentation in Real Orchard Environments Across Different Sensors [IROS2025]

Notifications You must be signed in to change notification settings

PRBonn/hapt3D

Repository files navigation

HAPT3D

Train

Run python train.py --config config/config_full.yaml. Remember to change the path to the dataset folder in the config file and in the train.py file.

Testing

Run python test.py -w <file>. Remember to change the path to the dataset folder in the config file and in the test.py file. If you want to test on the validation set, uncomment lines 41-44 in test.py.

Installation

After struggling a bit to install MinkowskiEngine, the procedure below is the one that worked out on my machine (operations to be done in that specific order):

    conda create --name hapt3d python=3.9
    conda activate hapt3d
    pip install torch==1.12.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113 --no-cache-dir
    pip install numpy==1.24.2
    pip install setuptools==60.0
    pip install pykeops --no-cache-di
    pip3 install -U git+https://github.com/NVIDIA/MinkowskiEngine -v --no-deps
    pip install pytorch-lightning==1.9.0 --no-deps
    pip install fsspec
    pip install lightning-utilities
    pip install tqdm
    pip install pyyaml
    pip install torchmetrics==1.4.1
    pip install ipdb
    pip install open3d
    pip install tensorboard
    pip install torchmetrics
    pip install hdbscan
    pip install distinctipy
    pip install optuna==3.6.1
    pip install optuna-integration

Good luck :)

Docker

Alternatively, you could simply use docker. Build it first via make build, then you can train via doing make train and test with make test CHECKPOINT=<file>.

About

3D Hierarchical Panoptic Segmentation in Real Orchard Environments Across Different Sensors [IROS2025]

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published