Unsupervised Multi-Agent and Single-Agent Perception from Cooperative Views

Project page for UMS, accepted at CVPR 2026.

Haochen Yang1, Baolu Li1, Lei Li1, Delin Ren1, Jiacheng Guo1, Minghai Qin1, Tianyun Zhang1, Hongkai Yu1
1Cleveland State University
Paper arXiv Code
UMS pipeline for unsupervised multi-agent and single-agent 3D perception

Abstract

UMS is an unsupervised framework that uses cooperative observations to learn both multi-agent and single-agent 3D object detection without human-annotated 3D boxes. Denser point clouds improve proposal classification, while geometric and semantic agreement across views supplies supervision to the single-agent branch.

Summary Video

Motivation

Benefits of cooperative views for point-cloud density and cross-view consensus

Method

UMS combines a Proposal Purifying Filter for removing unreliable candidate boxes, Progressive Proposal Stabilizing for easy-to-hard refinement, and Cross-View Consensus Learning for transferring multi-view geometric and semantic agreement to the single-agent detector.

Proposal Purifying Filter and Progressive Proposal Stabilizing modules

Results

UMS achieves the strongest unsupervised results on both V2V4Real and OPV2V. At IoU 0.5, it reaches 52.03 AP for multi-agent and 44.27 AP for single-agent perception on V2V4Real; on OPV2V, it reaches 83.89 and 71.30 AP, respectively.

UMS multi-agent and single-agent detection results on V2V4Real and OPV2V

Ablation Study

The core ablation isolates Proposal Purifying Filter, Progressive Proposal Stabilizing, and Cross-View Consensus Learning, showing how the components jointly improve multi-agent and single-agent perception.

Ablation study of each UMS component

Qualitative Results

V2V4Real

BibTeX

@inproceedings{yang2026ums,
  title     = {Unsupervised Multi-Agent and Single-Agent Perception from Cooperative Views},
  author    = {Yang, Haochen and Li, Baolu and Li, Lei and Ren, Delin and Guo, Jiacheng and Qin, Minghai and Zhang, Tianyun and Yu, Hongkai},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}