Facilitating Heterogeneous Sensor Information Cooperation in Multi-Agent Perception System

Published in Information Fusion on Apr.16, 2026.

Distributed Cooperation (DisCo) is a novel two-stage distributed paradigm. It aims to achieve robust multi-agent collaborative perception under arbitrary heterogeneous sensor configurations and dynamic addition/removal of agents.

Key Features

Contains unimodal checkpoints (each modality), which supports the following base encoding methods:

  • Camera: EfficientNet + LiftSplat
  • Camera: EfficientNet + Unprojection
  • LiDAR: PointPillars
  • LiDAR: SECOND

The checkpoints contain the following configurations of collaborative perception:

  • liftsplat (OPV2V), second (OPV2V), pillar (OPV2V&DAIR-V2X), pillar&liftsplat (OPV2V&DAIR-V2X), second&unprojection (DAIR-V2X): Base collaborative perception networks of homogeneous encodings

  • p + s / s + p (OPV2V): Collaborative perception networks of pillar and second encodings

  • s + pl / pl + s (OPV2V): Collaborative perception networks of second and pillar&liftsplat encodings

  • p + pl / pl + p (DAIR-V2X) : Collaborative perception networks of pillar and pillar&liftsplat encodings

  • pl + su / su + pl (DAIR-V2X): Collaborative perception networks of pillar&liftsplat and second&unprojection encodings

Due to differences in hardware architectures (e.g., different GPU types), CUDA/cuDNN versions, and floating-point precision, the numerical results may vary across different machines (maybe just slightly).

Citation

If you use DisCo and its checkpoints in your research, please use the following BibTeX entry.

@article{zhao2026facilitating,
  title={Facilitating heterogeneous sensor information cooperation in multi-agent perception system},
  author={Zhao, Binyu and Zhang, Wei and Zou, Zhaonian},
  journal={Information Sciences},
  pages={123529},
  year={2026},
  publisher={Elsevier}
}

Links

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading