Semi-Supervised Collaborative Perception via Masked Cross-View Reconstruction and Spatio-Temporal Consistency under Label Sparsity
Lists
Only you see this
Abstract
While collaborative 3D object detection has shown tremendous potential in intelligent transportation systems, its deployment is severely hindered by the excessive reliance on dense multi-agent annotations. The process of labeling dynamic 3D objects across multiple viewpoints is exceptionally resource-intensive. To mitigate this bottleneck, we investigate a highly practical yet challenging setting: collaborative perception with extremely limited annotations, approximating the absolute lower bound of supervised learning. To this end, we introduce MCR-ST, a unified semi-supervised learning framework designed to extract maximal utility from unlabeled multi-agent data. MCR-ST shifts the paradigm by integrating Augmentation Alignment with Spatio-Temporal Consistency. Specifically, the cross-view reconstruction module randomly masks feature patches from individual agents and reconstructs them using complementary views from neighboring vehicles, forcing the network to learn generalized, pose-invariant spatial representations without any manual labels. Concurrently, a spatio-temporal Multi-level Guidance consistency mechanism acts as a robust self-distillation guide, enforcing prediction alignment between Weak–Strong historical frames and the current collaborative view. This dual-pronged strategy effectively prevents the model from overfitting to the extremely sparse ground truths. Experimental results on standard collaborative benchmarks, including DAIR-V2X and OPV2V and WSAA, reveal that MCR-ST achieves remarkable detection accuracy in one-shot and two-shot scenarios, bridging the gap between fully supervised models and label-scarce cooperative driving systems.
BibTeX
@misc{chen2026semi,
author = {Chaoqi Chen},
title = {{Semi-Supervised Collaborative Perception via Masked Cross-View Reconstruction and Spatio-Temporal Consistency under Label Sparsity}},
year = {2026},
url = {https://openreview.net/forum?id=MBbhLplXL1},
howpublished = {OpenReview},
}Loading related content...