Reproduce “TriBand-BEV: Real-Time LiDAR-Only 3D Pedestrian Detection via Height-Aware BEV and High-Resolution Feature Fusion” 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
Mohammad Khoshkdahan, Alexey Vinel
Researchers at Karlsruhe Institute of Technology developed TriBand-BEV, a real-time LiDAR-only 3D object detection framework that uses a novel 3-channel height-aware BEV encoding and a high-resolution feature fusion network. The system achieves 49 frames per second and improves pedestrian BEV Average Precision by 7.76% (mean across difficulty) compared to Complex-YOLO on the KITTI dataset.
Key insights
- A compact 3-channel BEV representation effectively captures essential vertical structural cues by aggregating corrected reflectance across distinct height bands, enabling differentiation of objects.
- Preserving and fusing high-resolution feature maps in the network's neck is critical for improving the detection accuracy of small objects like pedestrians, which occupy minimal pixels in BEV.
- An IQR-based filter during 3D box recovery effectively handles noise and outliers in sparse LiDAR data, leading to more robust and geometrically plausible height estimations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.