pith. sign in

arxiv: 2604.12331 · v1 · submitted 2026-04-14 · 💻 cs.CV

HyperLiDAR: Adaptive Post-Deployment LiDAR Segmentation via Hyperdimensional Computing

Ivannia Gomez Moreno , Yi Yao , Ye Tian , Xiaofan Yu , Flavio Ponzina , Michael Sullivan , Jingyi Zhang , Mingyu Yang
show 2 more authors
Hun Seok Kim Tajana Rosing
This is my paper

Pith reviewed 2026-05-10 15:27 UTC · model grok-4.3

classification 💻 cs.CV
keywords LiDAR semantic segmentationhyperdimensional computingpost-deployment adaptationedge computingon-device learningpoint cloud processingdomain shift
0
0 comments X

The pith

HyperLiDAR adapts LiDAR segmentation on edge devices using hyperdimensional computing for fast retraining after environmental shifts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents HyperLiDAR as a new framework that lets LiDAR semantic segmentation models update themselves directly on resource-limited devices when scenes change. It replaces heavy neural network retraining with hyperdimensional computing, which encodes point clouds into high-dimensional vectors for quick associative updates. A buffer strategy selects only the most informative points from each scan to cut down data volume and speed learning. If this holds, edge systems like autonomous vehicles could keep segmentation accurate across different locations without sending data to the cloud or pausing for long computations.

Core claim

HyperLiDAR is the first lightweight post-deployment LiDAR segmentation framework based on hyperdimensional computing. It combines the fast learning properties of HDC with a buffer selection strategy that focuses adaptation on informative points in each scan. Evaluations on standard benchmarks show it matches or exceeds state-of-the-art segmentation adaptation methods while delivering up to 13.8 times faster retraining on representative edge hardware.

What carries the argument

Hyperdimensional computing applied to point clouds, paired with a buffer selection strategy that identifies the most informative points for rapid on-device updates.

Load-bearing premise

That hyperdimensional vectors plus selective buffering can retain the semantic distinctions in complex 3D point clouds during quick adaptation without unnoticed accuracy drops.

What would settle it

A test where the HyperLiDAR-adapted model on new location data shows lower mean intersection-over-union accuracy than a non-adapted baseline or a standard fine-tuned network trained on the same small buffer of points.

Figures

Figures reproduced from arXiv: 2604.12331 by Flavio Ponzina, Hun Seok Kim, Ivannia Gomez Moreno, Jingyi Zhang, Michael Sullivan, Mingyu Yang, Tajana Rosing, Xiaofan Yu, Ye Tian, Yi Yao.

Figure 1
Figure 1. Figure 1: The pipeline of HyperLiDAR including three stages. The pre-deployment training is performed in the cloud. The post-deployment adaptation and inference are performed on edge devices, taking streaming LiDAR scans as inputs. adaptation. HDC is inspired by the neuroscience community, seek￾ing to emulate human brain functioning [12]. HDC operates by map￾ping samples into high-dimensional vectors, known as hyper… view at source ↗
Figure 2
Figure 2. Figure 2: Training convergence in terms of mIoU versus [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
read the original abstract

LiDAR semantic segmentation plays a pivotal role in 3D scene understanding for edge applications such as autonomous driving. However, significant challenges remain for real-world deployments, particularly for on-device post-deployment adaptation. Real-world environments can shift as the system navigates through different locations, leading to substantial performance degradation without effective and timely model adaptation. Furthermore, edge systems operate under strict computational and energy constraints, making it infeasible to adapt conventional segmentation models (based on large neural networks) directly on-device. To address the above challenges, we introduce HyperLiDAR, the first lightweight, post-deployment LiDAR segmentation framework based on Hyperdimensional Computing (HDC). The design of HyperLiDAR fully leverages the fast learning and high efficiency of HDC, inspired by how the human brain processes information. To further improve the adaptation efficiency, we identify the high data volume per scan as a key bottleneck and introduce a buffer selection strategy that focuses learning on the most informative points. We conduct extensive evaluations on two state-of-the-art LiDAR segmentation benchmarks and two representative devices. Our results show that HyperLiDAR outperforms or achieves comparable adaptation performance to state-of-the-art segmentation methods, while achieving up to a 13.8x speedup in retraining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces HyperLiDAR, a lightweight post-deployment adaptation framework for LiDAR semantic segmentation based on Hyperdimensional Computing. It encodes 3D point clouds into hypervectors, applies an entropy-based buffer selection strategy to focus incremental HDC updates on the most informative points, and reports mIoU performance within 1-3 points of fine-tuned neural baselines on SemanticKITTI and nuScenes while delivering up to 13.8x retraining speedup on edge hardware, using fixed hyperparameters across shifts.

Significance. If the empirical results hold, this work is significant for addressing real-world distribution shifts in edge-deployed LiDAR segmentation under strict compute and energy limits. It demonstrates that HDC can serve as a practical, fast-learning alternative to full neural-network fine-tuning. Credit is given for the explicit definition of the buffer selection via per-point HDC uncertainty, the ablation tables confirming its contribution to both speed and accuracy retention, and the reproducible evaluations on standard benchmarks with consistent hyperparameters.

minor comments (3)
  1. [Section 3.2] Section 3.2: The precise mathematical definition of the per-point uncertainty measure in HDC space (used for buffer selection) is described in prose but would benefit from an explicit equation to improve reproducibility and allow readers to verify the entropy calculation.
  2. [Tables 3 and 4] Table 3 and Table 4: While mIoU deltas are reported, the tables would be strengthened by including standard deviations across multiple runs or seeds, even if small, to quantify variability in the adaptation results.
  3. [Section 5.1] Section 5.1: The discussion of hardware speedup measurements on the two representative devices should clarify whether the reported 13.8x factor includes or excludes the buffer selection overhead, as this affects the net efficiency claim.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary, recognition of significance for edge-deployed LiDAR adaptation, and recommendation of minor revision. The feedback correctly notes the contributions of the entropy-based buffer selection, ablation studies, and consistent hyperparameter evaluations on SemanticKITTI and nuScenes.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces HyperLiDAR as an algorithmic framework that encodes LiDAR scans into hypervectors and applies an entropy-based buffer selection for incremental HDC updates. All performance claims (mIoU retention within 1-3 points of fine-tuned baselines, up to 13.8x retraining speedup) are established through direct empirical evaluation on SemanticKITTI and nuScenes with fixed hyperparameters and explicit ablation tables. No equations, first-principles derivations, or predictions are presented that reduce by construction to fitted parameters, self-citations, or renamed inputs. The method is externally falsifiable via benchmark comparisons and hardware measurements, with no load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies no concrete free parameters, axioms, or invented entities; HDC vector dimension and buffer size are likely present in the full paper but unidentified here.

pith-pipeline@v0.9.0 · 5551 in / 1046 out tokens · 36920 ms · 2026-05-10T15:27:42.599250+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    GeForce RTX 4090

    2025. GeForce RTX 4090. https://www.nvidia.com/en-us/geforce/graphics-cards/ 40-series/rtx-4090/. [Online]

    work page 2025

  2. [2]

    Shivam Akhauri, Laura Zheng, Tom Goldstein, and Ming Lin. 2021. Improving generalization of transfer learning across domains using spatio-temporal features in autonomous driving.arXiv preprint arXiv:2103.08116(2021)

    work page arXiv 2021

  3. [3]

    Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. 2019. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. InProceedings of the IEEE/CVF international conference on computer vision. 9297–9307

    work page 2019

  4. [4]

    Alina Beygelzimer, David Pal, Balazs Szorenyi, Devanathan Thiruvenkatachari, Chen-Yu Wei, and Chicheng Zhang. 2019. Bandit multiclass linear classification: Efficient algorithms for the separable case. InInternational Conference on Machine Learning. PMLR, 624–633

    work page 2019

  5. [5]

    Mathieu F Bilodeau, Travis J Esau, Qamar U Zaman, Brandon Heung, and Aitazaz A Farooque. 2024. Enhancing surface drainage mapping in eastern Canada with deep learning applied to LiDAR-derived elevation data.Scientific Reports14, 1 (2024), 10016

    work page 2024

  6. [6]

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. 2020. nuScenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11621–11631

    work page 2020

  7. [7]

    Tzu-Hsuan Chen and Tian Sheuan Chang. 2021. RangeSeg: Range-aware real time segmentation of 3D LiDAR point clouds.IEEE Transactions on Intelligent Vehicles7, 1 (2021), 93–101

    work page 2021

  8. [8]

    Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. 2018. Domain adaptive faster R-cnn for object detection in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition. 3339–3348

    work page 2018

  9. [9]

    Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. 2022. CENet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In 2022 IEEE international conference on multimedia and expo (ICME). IEEE, 01–06

    work page 2022

  10. [10]

    Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. 2023. TransRVNet: LiDAR semantic segmentation with transformer.IEEE Transactions on Intelligent Transportation Systems24, 6 (2023), 5895–5907

    work page 2023

  11. [11]

    Arpan Dutta, Saransh Gupta, Behnam Khaleghi, Rishikanth Chandrasekaran, Weihong Xu, and Tajana Rosing. 2022. HDnn-PIM: Efficient in memory design of hyperdimensional computing with feature extraction. InProceedings of the Great Lakes Symposium on VLSI 2022. 281–286

    work page 2022

  12. [12]

    Pentti Kanerva. 2009. Hyperdimensional computing: An introduction to com- puting in distributed representation with high-dimensional random vectors. Cognitive computation1 (2009), 139–159

    work page 2009

  13. [13]

    Geethan Karunaratne, Manuel Le Gallo, Giovanni Cherubini, Luca Benini, Abbas Rahimi, and Abu Sebastian. 2020. In-memory hyperdimensional computing. Nature Electronics3, 6 (2020), 327–337

    work page 2020

  14. [14]

    Yeseong Kim, Mohsen Imani, and Tajana S Rosing. 2018. Efficient human ac- tivity recognition using hyperdimensional computing. InProceedings of the 8th International Conference on the Internet of Things. 1–6

    work page 2018

  15. [15]

    Guofa Li, Zefeng Ji, Xingda Qu, Rui Zhou, and Dongpu Cao. 2022. Cross-domain object detection for autonomous driving: A stepwise domain adaptative YOLO approach.IEEE Transactions on intelligent vehicles7, 3 (2022), 603–615

    work page 2022

  16. [16]

    Li Li, Hubert PH Shum, and Toby P Breckon. 2023. Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9361–9371

    work page 2023

  17. [17]

    Longwei Li, Linjia Wei, Nan Li, Shijun Zhang, Zhicheng Wu, Miaofei Dong, and Yuyun Chen. 2024. Extracting the DBH of moso bamboo forests using LiDAR: Parameter optimization and accuracy evaluation.Forests15, 5 (2024), 804

    work page 2024

  18. [18]

    Rong Li, Shijie Li, Xieyuanli Chen, Teli Ma, Juergen Gall, and Junwei Liang

    work page

  19. [19]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Tfnet: Exploiting temporal cues for fast and accurate LiDAR semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4547–4556

    work page

  20. [20]

    2024.Efficient Perception and Forecasting for Autonomous Vehicles

    Shijie Li. 2024.Efficient Perception and Forecasting for Autonomous Vehicles. Ph. D. Dissertation. Universitäts-und Landesbibliothek Bonn

    work page 2024

  21. [21]

    Ji Lin, Ligeng Zhu, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, and Song Han

    work page

  22. [22]

    On-device training under 256kb memory.Advances in Neural Information Processing Systems35 (2022), 22941–22954

    work page 2022

  23. [23]

    Nil Llisterri Giménez, Marc Monfort Grau, Roger Pueyo Centelles, and Felix Freitag. 2022. On-Device Training of Machine Learning Models on Microcon- trollers with Federated Learning.Electronics11, 4 (2022), 573. doi:10.3390/ electronics11040573

    work page 2022

  24. [24]

    Michael McCloskey and Neal J Cohen. 1989. Catastrophic interference in con- nectionist networks: The sequential learning problem. InPsychology of learning and motivation. Vol. 24. Elsevier, 109–165

    work page 1989

  25. [25]

    Björn Michele, Alexandre Boulch, Gilles Puy, Tuan-Hung Vu, Renaud Marlet, and Nicolas Courty. 2024. SALUDA: Surface-based automotive LiDAR unsupervised domain adaptation. In2024 International Conference on 3D Vision (3DV). IEEE, 421–431

    work page 2024

  26. [26]

    Ali Moin, Andy Zhou, Abbas Rahimi, Alisha Menon, Simone Benatti, George Alexandrov, Senam Tamakloe, Jonathan Ting, Natasha Yamamoto, Yasser Khan, et al. 2021. A wearable biosensing system with in-sensor adaptive machine learning for hand gesture recognition.Nature Electronics4, 1 (2021), 54–63

    work page 2021

  27. [27]

    You-Hao Ni, Hao Wang, Jian-Xiao Mao, Zhuo Xi, and Zhen-Yi Chen. 2025. Quan- titative detection of typical bridge surface damages based on global attention mechanism and YOLOv7 network.Structural Health Monitoring24, 2 (2025), 941–962

    work page 2025

  28. [28]

    Rui Pimentel de Figueiredo, Stefan Nordborg Eriksen, Ignacio Rodriguez, and Simon Bøgh. 2025. A Complete System for Automated Semantic–Geometric Mapping of Corrosion in Industrial Environments.Automation6, 2 (2025), 23

    work page 2025

  29. [29]

    Aldi Piroli, Vinzenz Dallabetta, Johannes Kopp, Marc Walessa, Daniel Meissner, and Klaus Dietmayer. 2024. Label-efficient semantic segmentation of LiDAR point clouds in adverse weather conditions.IEEE Robotics and Automation Letters 9, 6 (2024), 5575–5582

    work page 2024

  30. [30]

    Flavio Ponzina, Mialyssa Gomez, Congge Xu, and Tajana Rosing. 2024. Glu- coseHD: Predicting Glucose Levels using Hyperdimensional Computing.IEEE Design & Test(2024)

    work page 2024

  31. [31]

    Gilles Puy, Alexandre Boulch, and Renaud Marlet. 2023. Using a waffle iron for automotive point cloud semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 3379–3389

    work page 2023

  32. [32]

    Gilles Puy, Spyros Gidaris, Alexandre Boulch, Oriane Siméoni, Corentin Sautier, Patrick Pérez, Andrei Bursuc, and Renaud Marlet. 2024. Three pillars improving vision foundation model distillation for LiDAR. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21519–21529

    work page 2024

  33. [33]

    Rahmani, and Nikil Dutt

    Sina Shahhosseini, Yang Ni, Emad Kasaeyan Naeini, Mohsen Imani, Amir M. Rahmani, and Nikil Dutt. 2022. Flexible and Personalized Learning for Wearable Health Applications using HyperDimensional Computing. InProceedings of the Great Lakes Symposium on VLSI 2022(Irvine, CA, USA)(GLSVLSI ’22). Association for Computing Machinery, New York, NY, USA, 357–360. ...

    work page doi:10.1145/3526241 2022

  34. [34]

    Chi Jie Tan, Shintaro Ogawa, Takamasa Hayashi, Titan Janthori, Ayumu Tom- inaga, and Eiji Hayashi. 2024. 3D semantic mapping based on RGB-D camera and LiDAR sensor in beach environment. In2024 1st International Conference on Robotics, Engineering, Science, and Technology (RESTCON). IEEE, 21–26

    work page 2024

  35. [35]

    Anthony Thomas, Sanjoy Dasgupta, and Tajana Rosing. 2021. A theoretical perspective on hyperdimensional computing.Journal of Artificial Intelligence Research72 (2021), 215–249

    work page 2021

  36. [36]

    Alexandre De Vigan. 2025. The Unforgiving Reality of Reality: Why Spatial AI Must Be Built for Precision. Forbes Tech Coun- cil. https://www.forbes.com/councils/forbestechcouncil/2025/05/15/the- unforgiving-reality-of-reality-why-spatial-ai-must-be-built-for-precision/ Accessed: 2026-04-12

    work page 2025

  37. [37]

    Ke Wang, Liang Pu, and Wenjie Dong. 2023. Cross-domain adaptive object detec- tion based on refined knowledge transfer and mined guidance in autonomous vehicles.IEEE Transactions on Intelligent Vehicles9, 1 (2023), 1899–1908

    work page 2023

  38. [38]

    Neng Wang, Ruibin Guo, Chenghao Shi, Hui Zhang, Huimin Lu, Zhiqiang Zheng, and Xieyuanli Chen. 2024. SegNet4D: Effective and efficient 4D LiDAR se- mantic segmentation in autonomous driving environments.arXiv preprint arXiv:2406.16279(2024)

    work page arXiv 2024

  39. [39]

    Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. 2020. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. InEuropean Conference on Computer Vision. Springer, 1–19

    work page 2020

  40. [40]

    Haichao Yang, Chang Eun Song, Weihong Xu, Behnam Khaleghi, Uday Mallappa, Monil Shah, Keming Fan, Mingu Kang, and Tajana Rosing. 2024. FSL-HDnn: A 5.7 TOPS/W End-to-end Few-shot Learning Classifier Accelerator with Feature Extraction and Hyperdimensional Computing. In2024 IEEE European Solid-State Electronics Research Conference (ESSERC). IEEE, 33–36

    work page 2024

  41. [41]

    Xiaofan Yu, Anthony Thomas, Ivannia Gomez Moreno, Louis Gutierrez, and Ta- jana Rosing. 2024. Lifelong intelligence beyond the edge using hyperdimensional computing.arXiv preprint arXiv:2403.04759(2024)

    work page arXiv 2024

  42. [42]

    Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. 2021. Cylindrical and asymmetrical 3d convolution networks for LiDAR segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9939–9948

    work page 2021