Aligning Instance-Semantic Sparse Representation towards Unsupervised Object Segmentation and Shape Abstraction with Repeatable Primitives

This is an official PyTorch Implementation of Aligning Instance-Semantic Sparse Representation towards Unsupervised Object Segmentation and Shape Abstraction with Repeatable Primitives.

• PyTorch
• PyTorch3D
• SparseMax
• PyTorch Lightning
• tensorboard
• Open3D for visualization.

We use the ready-to-use datasets provised by CuboidAbstractionViaSeg:

Dataset

To start the training, run

    $ python train.py --data_dir /path/to/dataset

Please note that the sample_points_from_meshes function in PyTorch3D and the Sparsemax do not have a deterministic implementation. As a result, the following settings do not guarantee reproducibility:

os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
torch.use_deterministic_algorithms(True)
seed_everything(2024, workers=True)

The model is difficult to train. If retraining is needed, please check the validation results after 100 epochs. If the results do not meet expectations, restart the training.

To test a trained model, run

    $ python test.py --data_dir /path/to/dataset

We provide the trained weights and the final results:

Weights: ./trained/ShapeNetNormal4096.pth

Results on Baidu Netdisk Code: v799

Results on Google Drive

If you use our code for research, please cite our paper:

@ARTICLE{li2025aligning,
  author={Li, Jiaxin and Wang, Hongxing and Tan, Jiawei and Ou, Zhilong and Yuan, Junsong},
  journal={IEEE Transactions on Visualization and Computer Graphics}, 
  title={Aligning Instance-Semantic Sparse Representation towards Unsupervised Object Segmentation and Shape Abstraction with Repeatable Primitives}, 
  year={2025},
  volume={},
  number={},
  pages={},
  keywords={},
  doi={}
  }

Apache-2.0 Licence