QACE: Quality-Aware Collaborative Evidence for Semi-supervised Collaborative 3D Detection

Collaborative 3D perception can effectively enhance the perception capability of a single agent through information sharing, yet at the same time incurs expensive multi-agent joint annotation costs. Semi-supervised learning is a promising label-efficient solution, but it still faces two unique challenges in pseudo-labels, namely cross-agent observation inequality and fine-grained attribute disagreement; specifically, a missing proposal is informative only when the corresponding agent had sufficient opportunity to observe the object, and agreement on object presence does not imply equal reliability across box attributes. To address these issues, we propose QACE, a novel semi-supervised collaborative 3D detection framework that qualifies collaborative evidence before assigning pseudo-supervision. The Validated Collaborative Evidence (VCE) qualifies support, conflict, and silence using sensing opportunity and role-specific proposal checks. The Evidence-Conditioned Attribute Trainability (EAT) converts this evidence into separate weights for center, size, yaw axis, and directed heading within candidate-level supervision permissions. Extensive experiments on OPV2V and DAIR-V2X demonstrate that QACE achieves state-of-the-art performance across different detectors and annotation budgets, highlighting its effectiveness. The code and model will be publicly available.