arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Real‑time point‑BEV fusion for LiDAR semantic segmentation on edge devices. | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Computer VisionEnglishPublished

Real‑time point‑BEV fusion for LiDAR semantic segmentation on edge devices.

November 13, 2026arXiv: 2611.14025v1

HiPFuse is a framework that fuses individual point‑cloud data with a bird’s‑eye‑view (BEV) representation, letting a single LiDAR scan give both fine‑grained detail and broad 2D context. It exchanges features between points and height‑partitioned BEV grids at three horizontal scales while keeping the vertical groupings fixed, and it trains with an auxiliary loss that uses the same 3D cells to supervise initial BEV features.

On the widely used SemanticKITTI benchmark, HiPFuse reaches a mean intersection‑over‑union (mIoU) of 63.7 % while running at 14.2 frames per second on an NVIDIA Jetson AGX Orin. This speed exceeds the 10 Hz scan rate of the LiDAR, showing that the model can process a complete sweep faster than it is captured. The same architecture on the MCD benchmark achieves 62.5 % mIoU at 22.4 FPS.

Line plot of mean intersection‑over‑union (mIoU) versus frames per second on an NVIDIA Jetson AGX Orin, showing HiPFuse achieving 63.7 % mIoU at 14.2 FPS, above the 10 Hz benchmark threshold.
Fig. 1. SemanticKITTI test accuracy versus measured Jetson AGX Orin throughput (Tables III and V). Shading indicates throughput meeting the 10 Hz scan rate under the measured settings.
Wu, Haiyang, Vosselman, George, Lehtola, Ville · arXiv:2611.14025v1 · Fig. 1: · CC BY 4.0

The design means that replacing a conventional semantic front end in a robotic system can lower odometry error by about 10 % and raise accepted‑loop recall by almost nine percentage points on SemanticKITTI, while also building a higher quality semantic map. However, the spatial grid—its height range, cell size, and number of vertical partitions—is learned jointly with the network. Adapting a pre‑trained HiPFuse model to a different LiDAR or sensor setting therefore requires new labelled data and retraining.

Fig. 1 illustrates the accuracy–throughput trade‑off on SemanticKITTI, showing that HiPFuse consistently stays above the 10 Hz threshold under the benchmark’s measured Jetson AGX Orin settings. Fig. 2 visualises the HiPFuse pipeline, detailing how point features and BEV maps are exchanged across scales.

Schematic of HiPFuse architecture: LiDAR points are encoded, aggregated into height‑partitioned BEV grids, and features are exchanged between point and BEV branches across scales, then recombined for point‑wise semantic prediction.
Fig. 2. HiPFuse pipeline showing point‑to‑BEV feature exchange across three horizontal scales while maintaining fixed vertical partitions.Wu, Haiyang, Vosselman, George, Lehtola, Ville · arXiv:2611.14025v1 · Fig. 2: · CC BY 4.0

These results demonstrate that HiPFuse delivers high‑quality semantic segmentation on edge devices, laying groundwork for more accurate navigation and mapping in real‑time robotic applications.

Source paper

HiPFuse: Real-Time LiDAR Semantic Segmentation for Robotic Perception on Edge Devices

Computer VisionRobotics
View on arXiv