pith. sign in

arxiv: 2606.30937 · v1 · pith:Z4UYNVT4new · submitted 2026-06-29 · 💻 cs.CV

No Adaptation Without Observation: Observability-Constrained Test-Time Prompt Tuning for LiDAR Semantic Segmentation

Linlian Jiang , Wentao Ju , Sadman Rakib Pinon , Jianwei Xian , Zhixiang Chi , Xinxin Zuo , Yang Wang This is my paper

Pith reviewed 2026-07-01 01:39 UTC · model grok-4.3

classification 💻 cs.CV
keywords test-time adaptationLiDAR semantic segmentationprompt tuningpseudo-labelingobservabilitygeometry constraintsonline adaptation
0
0 comments X

The pith

A geometry-constrained prompt tuning method stabilizes test-time adaptation for LiDAR semantic segmentation by estimating per-location sensing reliability to gate updates.

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

The paper tackles degradation in LiDAR semantic segmentation when sensing conditions change at deployment, where new annotations are unavailable. Standard test-time adaptation injects unstable gradients because pseudo-label quality varies spatially with range-dependent sparsity and occlusion. The proposed framework confines updates to lightweight prompt adapters in a frozen backbone and reweights supervision using reliability estimates derived from depth-consistent beam terminations and neighborhood support. Spatial gating prevents unreliable regions from affecting shared parameters, while temporally smoothed prototype alignment accumulates reliable evidence over time. Experiments show this yields more stable adaptation and higher segmentation accuracy on standard benchmarks under deployment variations.

Core claim

By estimating per-location sensing reliability from depth-consistent beam terminations and neighborhood support, reweighting spatial supervision, confining adaptation to prompt adapters inserted into a frozen backbone with spatial gating to shield globally shared representations, and applying temporally smoothed prototype alignment, test-time adaptation for LiDAR semantic segmentation becomes stable and effective without additional annotations.

What carries the argument

Geometry-constrained test-time prompt tuning framework that computes per-location sensing reliability to reweight supervision and apply spatial gating on prompt adapter updates.

If this is right

  • Adaptation stays stable when sensing conditions evolve during online deployment.
  • Segmentation accuracy rises on standard LiDAR benchmarks that include deployment shifts.
  • Only lightweight prompt adapters receive updates while the backbone remains frozen.
  • Unreliable spatial regions are prevented from perturbing globally shared model parameters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same reliability estimation from beam terminations and neighborhood support could be tested on camera-based or radar-based segmentation where visibility also varies spatially.
  • Observability constraints of this form may generalize to other online pseudo-label settings that suffer from heteroscedastic noise, such as point cloud registration.
  • If the reliability map proves accurate, it could reduce reliance on periodic full retraining for fleets of LiDAR-equipped vehicles.

Load-bearing premise

That pseudo-label reliability can be reliably estimated from depth-consistent beam terminations and neighborhood support to safely gate which regions contribute to updates.

What would settle it

An ablation on the same LiDAR benchmarks that removes or randomizes the reliability-based reweighting and spatial gating, then measures whether adaptation stability and accuracy drop under the reported deployment variations.

Figures

Figures reproduced from arXiv: 2606.30937 by Jianwei Xian, Linlian Jiang, Sadman Rakib Pinon, Wentao Ju, Xinxin Zuo, Yang Wang, Zhixiang Chi.

Figure 1
Figure 1. Figure 1: We compute a geometry-aware observability score that [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Geometry-constrained test-time adaptation. Observability [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (A) Beam-termination and (B) neighborhood-supported [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on SemanticKITTI (Rows 1–3) and nuScenes (Rows 4–6). From left to right: RGB image, SFCNet [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of explicit observability modeling. (A) Ob [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

LiDAR semantic segmentation often degrades under real-world deployment due to evolving sensing conditions, while collecting new annotations for retraining is impractical. Test-time adaptation (TTA) updates model parameters online using pseudo-label supervision, but directly applying standard TTA strategies to LiDAR data is challenging. Because pseudo-label reliability is spatially heteroscedastic under range-dependent sparsity and occlusion, uniform updates on globally shared parameters can inject unstable gradients and destabilize adaptation. We propose a geometry-constrained test-time prompt tuning framework for LiDAR semantic segmentation. Our method estimates per-location sensing reliability from depth-consistent beam terminations and neighborhood support, and uses it to reweight spatial supervision. Adaptation is confined to lightweight prompt adapters inserted into a frozen backbone, with spatial gating to prevent unreliable regions from perturbing globally shared representations. A temporally smoothed prototype alignment strategy further stabilizes online updates by accumulating reliable semantic evidence over time. Experiments on standard LiDAR benchmarks demonstrate improved adaptation stability and segmentation performance under deployment variations without additional annotations.

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 / 0 minor

Summary. The manuscript proposes an observability-constrained test-time prompt tuning framework for LiDAR semantic segmentation. It estimates per-location sensing reliability from depth-consistent beam terminations and neighborhood support to reweight spatial supervision, confines adaptation to lightweight prompt adapters in a frozen backbone with spatial gating to avoid perturbing shared representations, and employs temporally smoothed prototype alignment for stability. Experiments on standard LiDAR benchmarks are reported to show improved adaptation stability and segmentation performance under deployment variations without additional annotations.

Significance. If the claimed improvements hold and the reliability estimator correlates with pseudo-label quality, the work would offer a practical, geometry-aware solution to a key deployment challenge in LiDAR segmentation by preventing unstable gradients from unreliable regions during online TTA. The prompt-tuning design is parameter-efficient and the spatial gating directly targets heteroscedastic reliability, which could generalize to other sparse sensing modalities.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for reviewing our manuscript and for the positive assessment of its potential significance in providing a geometry-aware solution for stable test-time adaptation in LiDAR semantic segmentation. We note that the recommendation is listed as uncertain, yet the report contains no specific major comments to address. We remain available to provide clarifications or additional experiments if requested.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper describes an algorithmic test-time adaptation method relying on geometric observations (depth-consistent beam terminations, neighborhood support) to estimate per-location reliability for spatial reweighting and gating of prompt adapters. No equations, fitted parameters, self-citations, or uniqueness theorems are referenced in the abstract or summary that would reduce any claimed prediction or result to its own inputs by construction. The approach is presented as a sequence of independent geometric constraints and stabilization strategies whose validity is assessed via external benchmark experiments rather than internal self-definition or renaming of known patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no free parameters, axioms, or invented entities are described or can be extracted.

pith-pipeline@v0.9.1-grok · 5724 in / 1095 out tokens · 31004 ms · 2026-07-01T01:39:28.526420+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

35 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    Rangevit: Towards vision transformers for 3d semantic segmentation in au- tonomous driving

    Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. Rangevit: Towards vision transformers for 3d semantic segmentation in au- tonomous driving. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 5240–5250, 2023. 1, 2, 5, 6, 7, 8

    2023

  2. [2]

    Se- mantickitti: A dataset for semantic scene understanding of lidar sequences

    Jens Behley, Martin Garbade, Andres Milioto, Jan Quen- zel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Se- mantickitti: A dataset for semantic scene understanding of lidar sequences. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 9297–9307,

  3. [3]

    Positioning and perception in lidar point clouds.Digital Signal Processing, 119:103193, 2021

    Csaba Benedek, Andras Majdik, Balazs Nagy, Zoltan Rozsa, and Tamas Sziranyi. Positioning and perception in lidar point clouds.Digital Signal Processing, 119:103193, 2021. 4

    2021

  4. [4]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 1, 5

    2020

  5. [5]

    Learning to adapt frozen clip for few-shot test-time domain adaptation

    Zhixiang Chi, Li Gu, Huan Liu, Ziqiang Wang, Yanan Wu, Yang Wang, and Konstantinos Plataniotis. Learning to adapt frozen clip for few-shot test-time domain adaptation. InIn- ternational Conference on Learning Representations, pages 66359–66380, 2025. 2

    2025

  6. [6]

    Octomap: An efficient probabilistic 3d mapping framework based on octrees.Au- tonomous robots, 34(3):189–206, 2013

    Armin Hornung, Kai M Wurm, Maren Bennewitz, Cyrill Stachniss, and Wolfram Burgard. Octomap: An efficient probabilistic 3d mapping framework based on octrees.Au- tonomous robots, 34(3):189–206, 2013. 3

    2013

  7. [7]

    Planning-oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17853–17862, 2023. 1

    2023

  8. [8]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. InEuropean conference on computer vision, pages 709–727. Springer, 2022. 2, 4, 7, 8

    2022

  9. [9]

    Spherical transformer for lidar-based 3d recognition

    Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 17545–17555, 2023. 1, 2, 5, 6

    2023

  10. [10]

    Temporal Ensembling for Semi-Supervised Learning

    Samuli Laine and Timo Aila. Temporal ensembling for semi- supervised learning.arXiv preprint arXiv:1610.02242, 2016. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  11. [11]

    Tcovis: Temporally consistent online video instance seg- mentation

    Junlong Li, Bingyao Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Tcovis: Temporally consistent online video instance seg- mentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1097–1107, 2023. 5

    2023

  12. [12]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in nat- ural language processing.ACM computing surveys, 55(9): 1–35, 2023

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hi- roaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in nat- ural language processing.ACM computing surveys, 55(9): 1–35, 2023. 2

    2023

  13. [13]

    P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks.arXiv preprint arXiv:2110.07602, 2021

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengx- iao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks.arXiv preprint arXiv:2110.07602, 2021. 2

    work page arXiv 2021

  14. [14]

    John Wiley & Sons, 2019

    Ricardo A Maronna, R Douglas Martin, Victor J Yohai, and Mat´ıas Salibi´an-Barrera.Robust statistics: theory and meth- ods (with R). John Wiley & Sons, 2019. 4

    2019

  15. [15]

    Rangenet++: Fast and accurate lidar semantic segmentation

    Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyrill Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In2019 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4213–4220. IEEE, 2019. 1, 2, 3, 5, 6

    2019

  16. [16]

    Exponential moving average of weights in deep learning: Dynamics and benefits.arXiv preprint arXiv:2411.18704, 2024

    Daniel Morales-Brotons, Thijs V ogels, and Hadrien Hen- drikx. Exponential moving average of weights in deep learning: Dynamics and benefits.arXiv preprint arXiv:2411.18704, 2024. 5

    work page arXiv 2024

  17. [17]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. InInterna- tional conference on machine learning, pages 16888–16905. PMLR, 2022. 2, 3

    2022

  18. [18]

    Fair-vpt: Fair visual prompt tuning for image classification

    Sungho Park and Hyeran Byun. Fair-vpt: Fair visual prompt tuning for image classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12268–12278, 2024. 2

    2024

  19. [19]

    Forecasting from lidar via future object detection

    Neehar Peri, Jonathon Luiten, Mengtian Li, Aljo ˇsa O ˇsep, Laura Leal-Taix ´e, and Deva Ramanan. Forecasting from lidar via future object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17202–17211, 2022. 1

    2022

  20. [20]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2

    2021

  21. [21]

    Gipso: Geometrically informed propa- gation for online adaptation in 3d lidar segmentation

    Cristiano Saltori, Evgeny Krivosheev, St ´ephane Lathuili´ere, Nicu Sebe, Fabio Galasso, Giuseppe Fiameni, Elisa Ricci, and Fabio Poiesi. Gipso: Geometrically informed propa- gation for online adaptation in 3d lidar segmentation. In European Conference on Computer Vision, pages 567–585. Springer, 2022. 3, 6, 7

    2022

  22. [22]

    Searching efficient 3d architec- tures with sparse point-voxel convolution

    Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architec- tures with sparse point-voxel convolution. InEuropean con- ference on computer vision, pages 685–702. Springer, 2020. 2

    2020

  23. [23]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017. 5

    2017

  24. [24]

    Tent: Fully Test-time Adaptation by Entropy Minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization.arXiv preprint arXiv:2006.10726,

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  25. [25]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7201–7211, 2022. 1, 3

    2022

  26. [26]

    Frnet: Frustum-range networks for scalable lidar segmenta- tion.IEEE Transactions on Image Processing, 2025

    Xiang Xu, Lingdong Kong, Hui Shuai, and Qingshan Liu. Frnet: Frustum-range networks for scalable lidar segmenta- tion.IEEE Transactions on Image Processing, 2025. 1, 2, 3, 5, 6, 7, 8

    2025

  27. [27]

    Prompt tuning for generative multimodal pretrained models.arXiv preprint arXiv:2208.02532, 2022

    Hao Yang, Junyang Lin, An Yang, Peng Wang, Chang Zhou, and Hongxia Yang. Prompt tuning for generative multimodal pretrained models.arXiv preprint arXiv:2208.02532, 2022. 2, 4

    work page arXiv 2022

  28. [28]

    Lidarmultinet: Towards a unified multi-task network for lidar perception

    Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 3231–3240, 2023. 1

    2023

  29. [29]

    Polarnet: An improved grid representation for online lidar point clouds se- mantic segmentation

    Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Ze- rong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds se- mantic segmentation. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 9601–9610, 2020. 1, 2, 5, 6

    2020

  30. [30]

    Spherical frustum sparse con- volution network for lidar point cloud semantic segmenta- tion.Advances in Neural Information Processing Systems, 37:121827–121858, 2024

    Yu Zheng, Guangming Wang, Jiuming Liu, Marc Polle- feys, and Hesheng Wang. Spherical frustum sparse con- volution network for lidar point cloud semantic segmenta- tion.Advances in Neural Information Processing Systems, 37:121827–121858, 2024. 1, 2, 5, 6, 7, 8

    2024

  31. [31]

    Conditional prompt learning for vision-language mod- els

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language mod- els. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 16816–16825,

  32. [32]

    Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

  33. [33]

    Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis

    Xin Zhou, Dingkang Liang, Wei Xu, Xingkui Zhu, Yihan Xu, Zhikang Zou, and Xiang Bai. Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14707– 14717, 2024. 2

    2024

  34. [34]

    Cylindrical and asymmetrical 3d convolution networks for lidar-based perception.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 44(10):6807–6822, 2021

    Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Wei Li, Yuexin Ma, Hongsheng Li, Ruigang Yang, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar-based perception.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 44(10):6807–6822, 2021. 1, 2, 5, 6

    2021

  35. [35]

    Hgl: Hierarchical geometry learning for test-time adaptation in 3d point cloud segmentation

    Tianpei Zou, Sanqing Qu, Zhijun Li, Alois Knoll, Lianghua He, Guang Chen, and Changjun Jiang. Hgl: Hierarchical geometry learning for test-time adaptation in 3d point cloud segmentation. InEuropean Conference on Computer Vision, pages 19–36. Springer, 2024. 3, 6, 7 10

    2024