Research Analyzer
← Back ICRA 2026

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

Yuheng Wu, Xiangbo Gao, Quang Tau, Zhengzhong Tu, Dongman Lee

PDF

AI summary

Key figure (auto-extracted from paper)
Background context is essential for robust collaborative perception and can be efficiently encapsulated into foreground features via curriculum learning, eliminating the need to transmit background data.
Collaborative perception foreground-centric background pruning curriculum learning bandwidth efficiency autonomous driving

Problem

Bandwidth constraints in vehicular networks make transmitting full feature maps impractical, while existing foreground-centric methods discard background regions despite their critical role in providing contextual cues for robust perception.

Approach

FadeLead uses a curricular training strategy that gradually prunes background features during learning, forcing the model to internalize essential context into compact foreground representations that are transmitted at inference.

Key results

  • Background regions encode essential contextual cues rather than redundant data
  • Curricular background pruning successfully internalizes context into foreground features
  • Outperforms prior methods across varying bandwidth constraints on simulated and real benchmarks
  • Maintains high detection accuracy even at extremely low feature transmission ratios (1%)

Why it matters

Enables bandwidth-efficient and robust collaborative perception for autonomous driving by proving that background context can be compressed into foreground features without increasing communication costs.

Abstract

Collaborative perception enhances the reliability and spatial coverage of autonomous vehicles by sharing com- plementary 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 imprac- tical. Recent methods, therefore, adopt a foreground-centric paradigm, transmitting only predicted foreground-region fea- tures while discarding the background, which encodes essential context. We propose FadeLead, a foreground-centric frame- work 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 pro- gressively 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. All artifacts are available at https://wyhallenwu.github. io/FadeLead/.

Index terms

Intelligent Transportation Systems Computer Vision for Transportation Cooperating Robots

Related papers