Reproduce “CBDES MoE: Hierarchically Decoupled Mixture-of-Experts for Functional Modules in Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qi Xiang, Kunsong Shi, Zhigui Lin, Lei He
Researchers at Tsinghua University and Imperial College London introduced CBDES MoE, a hierarchically decoupled Mixture-of-Experts framework for multi-modal Bird's Eye View (BEV) perception in autonomous driving. This system dynamically selects among heterogeneous expert networks, achieving a 1.6 mAP and 4.1 NDS improvement in 3D object detection on the nuScenes dataset while enhancing robustness in diverse conditions.
Key insights
- Leveraging architectural diversity among expert backbones significantly enhances scene adaptability and feature representation for robust autonomous driving.
- Dynamic, input-aware routing at the image level allows the system to adapt its computational path based on scene semantics, improving performance in varied conditions.
- Load balancing regularization is essential during training to ensure equitable utilization of all experts, preventing expert collapse and maximizing the model's overall capacity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.