HiPFuse: Real-Time LiDAR Semantic Segmentation for Robotic Perception on Edge Devices
HiPFuse: Real-Time LiDAR Semantic Segmentation for Robotic Perception on Edge Devices
This paper introduces HiPFuse, a point-BEV fusion framework designed for real-time LiDAR semantic segmentation on edge devices. The method exchanges point and height-partitioned BEV features across horizontal scales while preserving separate vertical aggregation groups through fixed height assignments. Evaluated on SemanticKITTI and MCD datasets, HiPFuse achieves 63.7% and 62.5% mIoU at 14.2 and 22.4 FPS respectively on Jetson AGX Orin, exceeding the 10 Hz scan rate. When integrated into fixed robotic systems, it reduces odometry absolute trajectory error RMSE by 10.6% on average across two outdoor sequences and increases accepted-loop recall by 8.7 percentage points on SemanticKITTI. The code is available at https://github.com/haiyang2022/HiPFuse.
The authors report HiPFuse as a point-BEV fusion framework achieving specific mIoU scores and FPS on Jetson AGX Orin hardware, with separate training/evaluation confirming indoor applicability and quantified improvements in odometry error and loop recall when replacing baseline semantic front ends.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
FPS on Jetson AGX Orin (SemanticKITTI)
14.2
View original evidence
t from one LiDAR scan. Fixed height assignments preserve separate vertical aggregation groups, with auxiliary supervision of initial BEV features using targets from the same 3D cells. A dual-scale BEV context module adds broader context from a coarser grid. On SemanticKITTI and MCD, HiPFuse achieves 63.7% and 62.5% mIoU at 14.2 and 22.4 FPS on Jetson AGX Orin, respectively, exceeding the 10 Hz scan rate. Separate tra
Open source S3FPS on Jetson AGX Orin (MCD)
22.4
View original evidence
t from one LiDAR scan. Fixed height assignments preserve separate vertical aggregation groups, with auxiliary supervision of initial BEV features using targets from the same 3D cells. A dual-scale BEV context module adds broader context from a coarser grid. On SemanticKITTI and MCD, HiPFuse achieves 63.7% and 62.5% mIoU at 14.2 and 22.4 FPS on Jetson AGX Orin, respectively, exceeding the 10 Hz scan rate. Separate tra
Open source S3
- code: Available Artifact link Open source S4
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
t from one LiDAR scan. Fixed height assignments preserve separate vertical aggregation groups, with auxiliary supervision of initial BEV features using targets from the same 3D cells. A dual-scale BEV context module adds broader context from a coarser grid. On SemanticKITTI and MCD, HiPFuse achieves 63.7% and 62.5% mIoU at 14.2 and 22.4 FPS on Jetson AGX Orin, respectively, exceeding the 10 Hz scan rate. Separate tra
Open source S3
ining and evaluation show indoor applicability. Replacing only the baseline semantic front end with HiPFuse in fixed robotic systems reduces odometry absolute trajectory error RMSE by 10.6% on average across two outdoor sequences and increases accepted-loop recall by 8.7 percentage points on SemanticKITTI. Qualitative semantic maps show more continuous road labeling. Our code is available at https://github.com/haiyan
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org