Multi-Tier Guidance for Multi-Agent 3D Detection under Severe Annotation Constraints
Lists
Only you see this
Abstract
Multi-agent collaborative perception boosts scene comprehension by allowing agents to exchange complementary sensor data. Yet, current collaborative 3D detection techniques depend heavily on extensive, fully labeled datasets, which are costly to acquire in multi-view settings due to alignment complexities and labeling consistency demands. This study explores collaborative 3D object detection in scenarios with drastically restricted annotations, where scenes have merely one or two ground-truth labels, establishing a minimal threshold for label scarcity. Such constraints severely hinder traditional supervised and semi-supervised approaches. To tackle this issue, we introduce a Strong-Weak Augmentation Matching with Multi-Tier Guidance framework tailored for collaborative 3D detection with minimal annotations. SWAMG incorporates two core modules. The first, Strong-Weak Augmentation Matching (SWAM), enforces consistency constraints between strongly and weakly augmented point clouds, facilitating robust utilization of vast unlabeled data in multi-agent environments. The second, Multi-Tier Guidance (MTG), employs a teacher network to deliver oversight at both label and feature tiers via pseudo-labeling and distillation, promoting enhanced alignment in representations between teacher and student. Through combined optimization of these consistency and guidance mechanisms, SWAMG mitigates overfitting and bolsters generalization amid extreme label limitations. Rigorous testing on OPV2V and DAIR-V2X benchmarks reveals that SWAMG attains state-of-the-art results in one-shot and two-shot configurations, markedly diminishing the need for voluminous annotated data in collaborative detection systems. These findings underscore the feasibility of developing dependable multi-agent perception models from highly sparse annotations.
BibTeX
@misc{luo2026multi,
author = {Bingjun Luo},
title = {{Multi-Tier Guidance for Multi-Agent 3D Detection under Severe Annotation Constraints}},
year = {2026},
url = {https://openreview.net/forum?id=7gg5FMUuMf},
howpublished = {OpenReview},
}Loading related content...