"ICME2025_HydraMamba_Supplementary"

"Abstract\u2014Current LiDAR-based 3D detection paradigms face a fundamental trade-off, sparse convolution is restricted bylocal receptive fields, while Transformer architectures incur prohibitive quadratic computational costs. Although State SpaceModels (SSMs) like Mamba offer linear complexity, their direct application to point clouds fractures local geometric topology andfails to address orientation-aware dependencies. We introduce HydraMamba, a hierarchical architecture that synthesizes theglobal modeling utility of Mamba with the local inductive bias of sparse convolution. Central to this architecture is the HydraBlock, which employs a Local-Global Gate to adaptively fuse features, complemented by Axial-Aware Aggregation and Content-Aware Merging to preserve structural integrity. Benchmarking on the KITTI dataset confirms HydraMamba establishes a newperformance standard with 85.2% mAP, surpassing Mamba baselines by 13.8% while maintaining real-time efficiency onconsumer hardware.This dataset contains supplementary materials for the paper: \"HydraMamba\" submitted to ICME 2026."