Reproduce “Lightweight 3D Object Detection via Mamba-Based Knowledge Distillation” 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
Quoc Cuong Ninh, Huy Xuan Pham, Anh Tung Nguyen, Dinh Hoan Trinh
A Mamba-based knowledge distillation framework produces lightweight 3D object detectors that maintain high detection accuracy while substantially reducing computational costs. The approach achieves 67.9% mAP and 71.9% NDS on nuScenes with 108 ms latency for a student model, outperforming existing methods on a real-world dataset and enabling deployment on embedded platforms at 3 Hz.
Key insights
- The multi-branch Mamba teacher architecture effectively combines shallow and deep Mamba branches to capture complementary multi-scale features and global context in 3D point clouds.
- Box-aware latent space distillation, focusing on object regions, significantly improves knowledge transfer for sparse 3D data compared to naive feature distillation across all voxels.
- Using a Mamba encoder as a projection module for latent space alignment proves effective in mapping student features to the teacher's feature space, enhancing distillation performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.