MHVTNet: Multi-Scale Hierarchical Voxel-Aware Transformer Network

MHVTNet is a 3D object detection framework designed for open-pit mining environments, addressing challenges such as sparse point clouds, weak and small targets, large target scale variations and dust-induced noise.

Data Preparation

If you want to train on AutoMine dataset, the data should be organized in the following KITTI-style structure:

data
│── automine
│   │── ImageSets/
│   │── training
│   │   │── label/
│   │   │── velodyne/
│   │   │── ...
│   │── testing
│   │   │── velodyne/
│   │   │── ...
│   │── gt_database/
│   │── automine_dbinfos_train.pkl
│   │── automine_infos_test.pkl
│   │── automine_infos_train.pkl
│   │── automine_infos_val.pkl
│   │── automine_infos_trainval.pkl

Training & Testing

To train and test MHVTNet, you need to import the project files into the corresponding folders of MMDetection3D. After that, you can follow the standard MMDetection3D workflow for training and testing.

Citation

@article{ZHANG2026130467,
    title = {Multi-scale hierarchical voxel-aware transformer network for 3D object detection in open-pit mines},
    journal = {Expert Systems with Applications},
    volume = {300},
    pages = {130467},
    year = {2026},
    issn = {0957-4174},
    doi = {https://doi.org/10.1016/j.eswa.2025.130467},
    url = {https://www.sciencedirect.com/science/article/pii/S0957417425040825},
    author = {Huazhen Zhang and Zhongyu Xie and Fan Zhang and Yuqian Zhao}
}

Acknowledgement

This project is mainly based on the following codebases. Thanks for their great works!

• MMDetection3D