Reproduce “SparseRadNet: Sparse Perception Neural Network on Subsampled Radar Data” on OpenResearch

About this paper

Jialong Wu, Mirko Meuter, Markus Schoeler, Matthias Rottmann

SparseRadNet introduces an efficient perception neural network that adaptively subsamples raw radar data, reducing it to 3% of original pixels for processing. This framework combines Graph Neural Networks and Sparse Convolutional Neural Networks, achieving a new state-of-the-art F1 score of 93.84 for object detection on the RADIal dataset with 10% fewer FLOPs compared to a dense baseline.

Key insights

• Adaptive, learned subsampling can effectively reduce radar data volume by up to 97% while retaining sufficient information for high-accuracy perception tasks.
• Combining GNNs and SCNNs in a two-branch architecture with attentive fusion provides a robust way to capture both global and local feature dependencies from sparse radar data.
• Incorporating radar-specific prior knowledge, such as object signature repetition handled by a sparse MIMO pre-encoder and axial attention for spatial alignment, is crucial for efficient and accurate sparse radar perception.

Summary by alphaXiv.