Background Fades, Foreground Leads: Curriculum-Guided Background Pruning for Efficient Foreground-Centric Collaborative Perception

TL;DR

This work tackles bandwidth-constrained collaborative perception by challenging the assumption that background information is redundant for robust multi-agent perception. It introduces FadeLead, a foreground-centric framework with three core components: Foreground Context Attention (FCA) to enrich foreground with scene context, Curricular Background Pruning (CBP) to progressively internalize background cues into compact foreground features during training, and Foreground Amplification Fusion (FAF) to robustly fuse ego and neighbor features while emphasizing foreground signals. Through a curriculum that starts with informative background and gradually prunes it, FadeLead achieves superior detection accuracy under tight bandwidth budgets across simulated and real-world datasets, and shows stability where reconstruction-based methods falter. The approach offers a practical path toward efficient, context-rich collaborative perception suitable for real-world vehicular networks, with notable gains at extremely low sharing ratios and scalable performance as bandwidth allows mild increases.

Abstract

Collaborative perception enhances the reliability and spatial coverage of autonomous vehicles by sharing complementary information across vehicles, offering a promising solution to long-tail scenarios that challenge single-vehicle perception. However, the bandwidth constraints of vehicular networks make transmitting the entire feature map impractical. Recent methods, therefore, adopt a foreground-centric paradigm, transmitting only predicted foreground-region features while discarding the background, which encodes essential context. We propose FadeLead, a foreground-centric framework that overcomes this limitation by learning to encapsulate background context into compact foreground features during training. At the core of our design is a curricular learning strategy that leverages background cues early on but progressively prunes them away, forcing the model to internalize context into foreground representations without transmitting background itself. Extensive experiments on both simulated and real-world benchmarks show that FadeLead outperforms prior methods under different bandwidth settings, underscoring the effectiveness of context-enriched foreground sharing.

Figures (8)

• Figure 1: (a) Average Precision (AP) of Where2Comm huWhere2commCommunicationEfficientCollaborative2022 on OPV2V opv2v under different communication strategies and (b) visualization of the post-fusion feature map.
• Figure 2: Comparison of different foreground-centric sharing strategy.
• Figure 3: Overview of FadeLead.
• Figure 4: Visualization on OPV2V opv2v with top-1% confident foreground selection. We show (a) detection results, (b) the ego BEV feature before fusion, (c) the ego BEV feature after fusion, and (d) the shared regions from collaborating CAVs.
• Figure 5: Visualization of detection results, post-fusion features, and shared FG regions under different top-$k$ ratios on V2X-R V2X-R.