Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Domain Adaptation for Different Sensor Configurations in 3D Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Two-stage training with partial layer freezing adapts 3D object detectors across different LiDAR sensor configurations, beating naive joint training on both taxi and bus datasets.

desk verdict Useful new benchmark and a plausible recipe, but the paper contradicts its own sensor specs to justify the method, and the numbers are thin. read the letter →

arxiv 2509.04711 v1 pith:W2PZ5SWP submitted 2025-09-04 cs.CV cs.RO

Satoshi Tanaka , Kok Seang Tan , Isamu Yamashita This is my paper · ORCID
classification cs.CVcs.RO
keywords 3DobjectdetectionLiDARdomainadaptationsensorconfigurationfine-tuningmulti-datasettrainingautonomousdrivingpointcloud
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tackles a practical gap in 3D object detection: vehicles on the same fleet often carry different LiDAR layouts, so a model trained on one configuration loses accuracy on another even in the same geographic region. Using paired datasets from a taxi-like single-top-LiDAR setup and a bus-like multi-corner-LiDAR setup with identical annotation classes, the authors show that training jointly on both configurations and then fine-tuning each configuration separately beats naive joint training. They also show the fine-tuning should be partial: retraining only the backbone and neck while freezing the encoder and head gives the best results, with mAP rising from 63.3 to 64.7 on the taxi configuration and from 64.5 to 65.6 on the bus configuration. The practical upshot is a scalable recipe for deploying one base model across heterogeneous vehicle platforms without sacrificing per-configuration accuracy.

What carries the argument

The carrying mechanism is a two-stage training pipeline. Stage one is joint training on all sensor configurations to learn shared, generalizable representations; stage two is Downstream Fine-tuning per configuration, restricted by Partial Layer Fine-tuning to the backbone and neck modules while the encoder and head are frozen. The rationale is module-level: the encoder sees similar concatenated point-cloud inputs, the head predicts from a shared annotation space, and only the backbone and neck spatial-feature extraction is sensitive to sensor placement. The paired RoboTaxi/RoboBus dataset with unified annotations supplies the controlled testbed.

What would settle it

Swap in a second annotation schema for one configuration, such as merging or dropping a class while keeping the sensors identical; if freezing the head then lowers mAP relative to unfreezing it, the shared-output assumption underlying the recipe breaks.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the sensor-configuration domain gap lives mostly in mid-level spatial features, not in the input encoder or the output head. Because the two configurations share the same annotation scheme and similar mechanical-LiDAR input statistics, the paper argues the encoder and head can be frozen, while the backbone and neck, which aggregate spatial features, must be retrained. Carried through Downstream Fine-tuning and Partial Layer Fine-tuning, this yields consistent gains over naive joint training on both configurations, with the best partial-freeze choice being the backbone and neck only. The companion unsupervised experiment indicates that t

Load-bearing premise

The recipe assumes the two configurations share the same annotation space and similar input LiDAR statistics, so freezing the encoder and head is safe; if a new platform changes the object classes or the LiDAR type, that premise gives way.

Editorial extensions

If this is right

  • Joint training alone already beats single-configuration training on both targets, but the proposed downstream fine-tuning adds another one to two mAP points per configuration.
  • Freezing the encoder and head while training only the backbone and neck outperforms full fine-tuning, so configuration adaptation can be achieved with fewer updated parameters.
  • Fine-tuning from a single-configuration model with 100% of the target data matches joint-training performance, and even 10% of the target data closes most of the cross-configuration gap.
  • The unsupervised pseudo-label experiment shows only marginal gains, supporting the paper's characterization of the sensor-configuration gap as a refinement problem rather than a detection problem.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the neck really is the most sensor-sensitive module, as the supplemental relative parameter differences suggest, a practical extension would be a cheap probe run that decides which layers to unfreeze from L1 parameter shifts, replacing the hand-designed split.
  • The same two-stage recipe should be tested on other detector architectures and on sensor pairs with very different beam counts or solid-state LiDAR; the paper demonstrates it only on a CenterPoint detector over one taxi/bus pair.
  • A targeted experiment could test the head-freezing assumption directly: keep the sensors identical but change the annotation schema, and if freezing the head then hurts, the method's scope is narrower than 'sensor configurations'.
  • A testable extension for the unsupervised setting is to generate pseudo-labels from the joint model with uncertainty filtering for small objects, which the paper identifies as the main bottleneck in cross-configuration pseudo-label adaptation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper studies 3D object detection under a sensor-configuration domain gap. The authors construct paired RoboTaxi and RoboBus datasets collected in Japan under a unified annotation format but with substantially different LiDAR layouts. They propose two techniques: Downstream Fine-tuning (jointly train on all configurations, then fine-tune per configuration) and Partial Layer Fine-tuning (freeze certain modules and update only a subset). Using CenterPoint, they report that joint training improves over single-dataset baselines, downstream fine-tuning adds a further 1-2 mAP, and the best partial-freeze scheme updates only the backbone and neck. An unsupervised pseudo-label experiment is also reported. The central empirical claim is that the proposed multi-stage fine-tuning strategy consistently outperforms naive joint training for each sensor configuration.

Significance. If the empirical claims hold, the paper offers a simple and practical recipe for adapting a 3D detector to heterogeneous vehicle platforms. The main contribution is the paired multi-configuration dataset: two sensor layouts in the same geographic region with identical annotation semantics, which isolates the sensor-induced domain gap from environmental and labeling differences. This is a valuable resource for the community. The proposed methods are straightforward and likely easy to reproduce with standard libraries. The analysis also includes a useful ablation over which modules to freeze. However, the broader significance is weakened by the lack of statistical validation and by an internal inconsistency in the justification for the freezing scheme.

major comments (3)
  1. [§3.2 vs §3.1] The justification for freezing the encoder and head is internally contradicted by the stated sensor specifications. §3.2 says 'Since the same mechanical LiDAR is used across configurations and both utilize concatenated point clouds, the encoder experiences similar input distributions and is therefore kept fixed.' But §3.1 describes RoboTaxi as a 128-beam top LiDAR plus 16-beam sub-LiDARs, and RoboBus as multiple 40-beam main LiDARs plus 64-beam short-range LiDARs. These produce different point densities and intensity patterns, so the premise of similar input distributions is not supported. The ablation in Table 4 may still empirically favor freezing these modules, but the proposed mechanism in §3.2 is not a valid explanation. Please correct the factual description or provide a principled argument that the concatenated input distributions are equivalent despite the different beam counts.
  2. [§4.2, Tables 3 and 4] All results appear to be single runs with no error bars, repeated seeds, or significance tests. The headline gains are small: 64.7 vs 63.3 mAP on RoboTaxi and 65.6 vs 64.5 on RoboBus. In the ablation, the best backbone+neck configuration beats several alternatives by only 0.1-0.4 mAP (e.g., 64.7 vs 64.6 vs 64.2 in Table 4a; 65.6 vs 65.3 vs 65.2 in Table 4b). Without variance estimates, these differences are within the range that could arise from random seed variation, and the claim that backbone+neck freezing is 'best' is not statistically supported. Please add at least multiple seeds with standard deviations or a paired significance test, and temper the wording accordingly.
  3. [Appendix §7.2-§7.3 and §3.2] The choice of which layers to freeze appears to be derived from a parameter-difference analysis performed on the same data that is later used for evaluation. The Appendix reports comparing 'the RoboTaxi model and the RoboBus model' and 'the base model and the domain-specific models' to determine which modules are most affected by domain shift. If these analyses are done on the same train/val/test partitions used in Table 4, then selecting the freeze scheme based on those figures is a form of validation-set selection and the reported performance may overfit this choice. Please clarify whether a separate held-out split or cross-validation was used, or otherwise present the layer choice as post-hoc analysis rather than as a prescriptive method.
minor comments (6)
  1. [§4.1] The mAP definition is said to follow 'the same mAP definition as the nuScenes benchmark.' For reproducibility, please specify the exact distance thresholds, matching criteria, and whether the reported mAP is averaged over distance ranges or classes.
  2. [Table 3 caption] The abbreviations 'RT.', 'RB.', 'FT.', 'Tru.', 'Bic.', 'Ped.' are not all defined in the caption. Please expand them or add a legend.
  3. [References] Several LLM-related references [1,2,4,17,19,27,44,52] are cited only in the supplementary section on LLM fine-tuning, not in the main text. The connection to LLMs is tangential to the paper's core contribution; consider removing or heavily condensing that supplementary discussion.
  4. [§2] The sentence 'However, as summarized by the experimental results of MDT3D [28] using CenterPoint in Table 1, show that...' has a grammatical error: 'show that' should be 'shows that' or the sentence restructured.
  5. [Various] There are typographical inconsistencies such as 'V oxel' in §4.1 and missing spaces in Algorithm 1 ('is relative=TrueandD sum[l]>0'). Please proofread.
  6. [§4.3/Table 6] The unsupervised experiment reports a marginal improvement from 52.0 to 52.3 mAP and attributes the difficulty to small objects. It would be helpful to also report results using the same fine-tuning strategy from a joint-training base model, to separate the effect of the base model from the pseudo-label quality.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical comparisons of fine-tuning protocols; no prediction reduces by construction to a fitted constant or to a self-citation.

full rationale

The paper's central claims are empirical: joint training plus Downstream Fine-tuning and Partial Layer Fine-tuning is evaluated on a newly collected paired dataset. There is no equation-level derivation in which an output quantity equals an input by construction. Downstream Fine-tuning is a standard two-stage training protocol; its reported gains are measured on held-out test splits, not derived from the method's definition. Partial Layer Fine-tuning's choice of freezing encoder/head and updating backbone/neck is justified by qualitative reasoning about input distributions and output annotation spaces (Section 3.2) and by a post-hoc layer-wise parameter-difference analysis (Appendix 7). This is model selection / exploratory analysis rather than a fitted parameter renamed as a prediction; the Table 4 ablation is an independent empirical comparison. There are no load-bearing self-citations or imported uniqueness theorems. One internal inconsistency exists: Section 3.2 says 'the same mechanical LiDAR is used across configurations' while Section 3.1 describes RoboTaxi with 128/16-beam LiDARs and RoboBus with 40/64-beam LiDARs. This weakens the stated mechanism for freezing the encoder, but it is a factual/consistency concern, not circularity. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of domain assumptions about the representativeness of the two sensor configurations, the separability of the sensor-configuration gap from other dataset differences, and the validity of freezing the encoder and head. No free parameters are fitted; the only hand choice is which layers to update, which is validated by the ablation. No invented entities are introduced.

assumptions (4)
  • domain assumption The RoboTaxi and RoboBus configurations represent the real-world sensor-placement domain gap, and identical annotation plus same geographic region isolates that gap.
    Section 3.1 constructs the dataset and frames it as representative of taxi and bus platforms; the entire analysis attributes measured mAP differences to sensor configuration, though the two setups also differ in beam counts and LiDAR models.
  • domain assumption The encoder and head can be frozen because input point-cloud distributions and output annotation formats are shared across configurations.
    Section 3.2 ('the encoder experiences similar input distributions and is therefore kept fixed', 'the head is also fixed') is load-bearing for Partial Layer Fine-tuning; it is checked only on two configurations in the Table 4 ablation.
  • domain assumption nuScenes-style mAP is a valid performance measure for comparing models across these configurations.
    Section 4.1 states the evaluation follows the nuScenes mAP definition; the choice affects all reported conclusions.
  • domain assumption Standard supervised training and the chosen hyperparameters (voxel 0.32 m, range 120 m, no multi-frame) do not confound the comparison.
    Section 4.1 fixes these for all experiments, but there is no sensitivity analysis; the paper does not report learning rate, epochs, batch size, or augmentation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain Adaptation for Different Sensor Configurations in 3D Object Detection." pith.science (2026). https://pith.science/paper/W2PZ5SWP

@misc{pith2026250904711,
  author       = {Pith},
  title        = {Pith review of: Domain Adaptation for Different Sensor Configurations in 3D Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W2PZ5SWP}},
  note         = {Machine review of arXiv:2509.04711}
}
read the original abstract

Recent advances in autonomous driving have underscored the importance of accurate 3D object detection, with LiDAR playing a central role due to its robustness under diverse visibility conditions. However, different vehicle platforms often deploy distinct sensor configurations, causing performance degradation when models trained on one configuration are applied to another because of shifts in the point cloud distribution. Prior work on multi-dataset training and domain adaptation for 3D object detection has largely addressed environmental domain gaps and density variation within a single LiDAR; in contrast, the domain gap for different sensor configurations remains largely unexplored. In this work, we address domain adaptation across different sensor configurations in 3D object detection. We propose two techniques: Downstream Fine-tuning (dataset-specific fine-tuning after multi-dataset training) and Partial Layer Fine-tuning (updating only a subset of layers to improve cross-configuration generalization). Using paired datasets collected in the same geographic region with multiple sensor configurations, we show that joint training with Downstream Fine-tuning and Partial Layer Fine-tuning consistently outperforms naive joint training for each configuration. Our findings provide a practical and scalable solution for adapting 3D object detection models to the diverse vehicle platforms.

Figures

Figures reproduced from arXiv: 2509.04711 by the authors.

Figure 1
Figure 1. Overview of the proposed training strategy combining Downstream Fine-tuning and Partial Layer Fine-tuning, as our concept [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Multi-configuration dataset with a consistent annotation format across different sensor setups. The dataset includes RoboTaxi [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The total number of parameters for each layer. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The absolute parameter differences between the Robo [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: The relative parameter differences between the Robo [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A User-driven Design Framework for Robotaxi

    cs.RO 2026-02 conditional novelty 5.0 of 10

    Real-world Chinese robotaxi users value driverless privacy and consistent driving but struggle with rigid stops, poor transparency, and unclear emergency controls; these findings support a four-phase, user-driven desi...

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [1]

    Llama 2: Open foundation and fine- tuned chat models, 2023

    Llama Team at Meta. Llama 2: Open foundation and fine- tuned chat models, 2023. 1

    work page 2023

  2. [2]

    The llama 3 herd of models, 2024

    Llama Team at Meta. The llama 3 herd of models, 2024. 1

    work page 2024

  3. [3]

    Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transform- ers.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1080–1089, 2022. 2, 4

    work page 2022

  4. [4]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz...

    2020

  5. [5]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom

    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. InCVPR, 2020. 2

    2020

  6. [6]

    Wellington

    Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi- Gonzalez, and Carl K. Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map cre- ation, localization, and perception.IEEE Signal Processing Magazine, 38:68–86, 2020. 1

    work page 2020

  7. [7]

    Cmd: A cross mechanism domain adaptation dataset for 3d object detection

    Jinhao Deng, Wei Ye, Hai Wu, Xun Huang, Qiming Xia, Xin Li, Jin Fang, Wei Li, Chenglu Wen, and Cheng Wang. Cmd: A cross mechanism domain adaptation dataset for 3d object detection. InComputer Vision – ECCV 2024, pages 219–

    work page 2024

  8. [8]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. InConference on Computer Vision and Pattern Recog- nition (CVPR), 2012. 2

    work page 2012

Show all 57 references
  1. [9]

    Density-insensitive unsupervised domain adaption on 3d object detection

    Qianjiang Hu, Daizong Liu, and Wei Hu. Density-insensitive unsupervised domain adaption on 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2023. 3

    2023

  2. [10]

    Semi- supervised domain adaptation using target-oriented domain augmentation for 3d object detection.IEEE Transactions on Intelligent Vehicles, pages 1–12, 2024

    Yecheol Kim, Junho Lee, Changsoo Park, Hyoung won Kim, Inho Lim, Christopher Chang, and Jun Won Choi. Semi- supervised domain adaptation using target-oriented domain augmentation for 3d object detection.IEEE Transactions on Intelligent Vehicles, pages 1–12, 2024. 3

    2024

  3. [11]

    Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom

    Alex H. Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689–12697, 2018. 2, 4

    2019

  4. [12]

    Domain general- ization of 3d object detection by density-resampling

    Shuangzhi Li, Lei Ma, and Xingyu Li. Domain general- ization of 3d object detection by density-resampling. In European Conference on Computer Vision, pages 456–473. Springer, 2024. 3

    2024

  5. [13]

    Dali: Domain adaptive li- dar object detection via distribution-level and instance-level pseudo label denoising.IEEE Transactions on Robotics,

    Xiaohu Lu and Hayder Radha. Dali: Domain adaptive li- dar object detection via distribution-level and instance-level pseudo label denoising.IEEE Transactions on Robotics,

  6. [14]

    One million scenes for autonomous driv- ing: Once dataset

    Jiageng Mao, Minzhe Niu, Chenhan Jiang, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, et al. One million scenes for autonomous driv- ing: Once dataset. 2021. 3

    2021

  7. [15]

    3d object detection for autonomous driving: A comprehensive survey, 2023

    Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hong- sheng Li. 3d object detection for autonomous driving: A comprehensive survey, 2023. 1

    2023

  8. [16]

    Meituan: Autonomous delivery and urban ser- vice platform.https://about.meituan.com/en,

    Meituan Inc. Meituan: Autonomous delivery and urban ser- vice platform.https://about.meituan.com/en,

  9. [17]

    Large language models: A survey, 2025

    Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey, 2025. 1

    2025

  10. [18]

    Nuro: Autonomous delivery for everyday needs

    Nuro, Inc. Nuro: Autonomous delivery for everyday needs. https://www.nuro.ai/, 2025. Accessed: 2025-05-

    2025

  11. [19]

    Gpt-4 technical report, 2024

    OpenAI. Gpt-4 technical report, 2024. 1

    2024

  12. [20]

    Cl3d: Unsuper- vised domain adaptation for cross-lidar 3d detection

    Xidong Peng, Xinge Zhu, and Yuexin Ma. Cl3d: Unsuper- vised domain adaptation for cross-lidar 3d detection. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 2047–2055, 2023. 3

  13. [21]

    Plus: Transforming the trucking industry with self- driving technology.https://www.plus.ai/, 2025

    Plus, Inc. Plus: Transforming the trucking industry with self- driving technology.https://www.plus.ai/, 2025. Accessed: 2025-05-10. 1

    2025

  14. [22]

    Pony.ai: Autonomous driving technology

    Pony.ai, Inc. Pony.ai: Autonomous driving technology. https://www.pony.ai/, 2025. Accessed: 2025-05-

    2025

  15. [23]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point- net: Deep learning on point sets for 3d classification and seg- mentation.2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77–85, 2016. 2

    2017

  16. [24]

    Qi, Li Yi, Hao Su, and Leonidas J

    Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point- net++: deep hierarchical feature learning on point sets in a metric space. InProceedings of the 31st International Conference on Neural Information Processing Systems, page 5105–5114, Red Hook, NY , USA, 2017. Curran...

    2017

  17. [25]

    Pointr- cnn: 3d object proposal generation and detection from point cloud.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–779, 2018

    Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr- cnn: 3d object proposal generation and detection from point cloud.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–779, 2018

    2019

  18. [26]

    Pv-rcnn: Point-voxel feature set abstraction for 3d object detection

    Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jian- ping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 10526–10535, 2019. 2

    2020

  19. [27]

    Karan Singhal, Shekoofeh Azizi, Tao Tu, S. Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tan- wani, Heather Cole-Lewis, Stephen Pfohl, Perry Payne, Mar- tin Seneviratne, Paul Gamble, Chris Kelly, Abubakr Babiker, Nathanael Sch¨arli, Aakanksha Chowdhery, Philip ...

    2023

  20. [28]

    Mdt3d: Multi-dataset training for lidar 3d object detection generalization

    Louis Soum-Fontez, Jean-Emmanuel Deschaud, and Franc ¸ois Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5765–5772, 2023. 2, 3

    2023

  21. [29]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Et- tinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhan...

    2020

  22. [30]

    Triess, Mariella Dreissig, Christoph B

    Larissa T. Triess, Mariella Dreissig, Christoph B. Rist, and Johann Marius Z¨ollner. A survey on deep domain adaptation for lidar perception.2021 IEEE Intelligent Vehicles Sympo- sium Workshops (IV Workshops), pages 350–357, 2021. 1

    2021

  23. [31]

    Ms3d++: Ensemble of experts for multi-source unsupervised domain adaptation in 3d ob- ject detection.IEEE Transactions on Intelligent Vehicles, pages 1–16, 2024

    Darren Tsai, Julie Stephany Berrio, Mao Shan, Eduardo Nebot, and Stewart Worrall. Ms3d++: Ensemble of experts for multi-source unsupervised domain adaptation in 3d ob- ject detection.IEEE Transactions on Intelligent Vehicles, pages 1–16, 2024. 3

    2024

  24. [32]

    Tusimple: Self-driving technology for long- haul freight.https://www.tusimple.com/, 2025

    TuSimple, Inc. Tusimple: Self-driving technology for long- haul freight.https://www.tusimple.com/, 2025. Accessed: 2025-05-10. 1

    2025

  25. [33]

    Weinberger, and Wei- Lun Chao

    Yan Wang, Xiangyu Chen, Yurong You, Li Erran, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger, and Wei- Lun Chao. Train in germany, test in the usa: Making 3d ob- ject detectors generalize. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2020

  26. [34]

    Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud

    Yan Wang, Junbo Yin, Wei Li, Pascal Frossard, Ruigang Yang, and Jianbing Shen. Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. InPro- ceedings of the AAAI Conference on Artificial Intelligence,

  27. [35]

    One for all: Multi-domain joint training for point cloud based 3d object detection

    Zhenyu Wang, Yali Li, Hengshuang Zhao, and Shengjin Wang. One for all: Multi-domain joint training for point cloud based 3d object detection. InAdvances in Neural In- formation Processing Systems, 2024. 3

    2024

  28. [36]

    Waymo: Autonomous driving technology company.https://waymo.com/, 2025

    Waymo LLC. Waymo: Autonomous driving technology company.https://waymo.com/, 2025. Accessed: 2025-05-09. 1

    2025

  29. [37]

    Weride - leading autonomous driving company

    WeRide Inc. Weride - leading autonomous driving company. https://www.weride.ai/, 2025. Accessed: 2025- 05-10. 1

    2025

  30. [38]

    Wozniak, Mattias Hansson, Marko Thiel, and Patric Jensfelt

    Maciej K. Wozniak, Mattias Hansson, Marko Thiel, and Patric Jensfelt. Uada3d: Unsupervised adversarial domain adaptation for 3d object detection with sparse lidar and large domain gaps.arXiv preprint arXiv:2403.17633, 2024. 3

    2024 arXiv

  31. [39]

    Towards universal lidar-based 3d object detec- tion by multi-domain knowledge transfer

    Guile Wu, Tongtong Cao, Bingbing Liu, Xingxin Chen, and Yuan Ren. Towards universal lidar-based 3d object detec- tion by multi-domain knowledge transfer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8669–8678, 2023. 3

    2023

  32. [40]

    Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudo-labeling

    Chao Xu, Chenghao Liu, Yuwei Liu, Yuntao Li, Xinyu Chen, Zhi Zhou, Junchi Zhang, Ming Yang, Qiang Yu, and Hao Tang. Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudo-labeling. In Proceedings of the IEEE/CVF International Conference o...

    2023

  33. [41]

    Qi, and Dragomir Anguelov

    Qiangeng Xu, Yin Zhou, Weiyue Wang, Charles R. Qi, and Dragomir Anguelov. Spg: Unsupervised domain adaptation for 3d object detection via semantic point generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15446–15456, 2021. 3

    2021

  34. [42]

    Second: Sparsely embed- ded convolutional detection.Sensors (Basel, Switzerland), 18, 2018

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embed- ded convolutional detection.Sensors (Basel, Switzerland), 18, 2018. 2, 4

    2018

  35. [43]

    Pixor: Real- time 3d object detection from point clouds.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652–7660, 2018

    Binh Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Real- time 3d object detection from point clouds.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652–7660, 2018. 2

    2018

  36. [44]

    Fingpt: Open-source financial large language models, 2023

    Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. Fingpt: Open-source financial large language models, 2023. 1

    2023

  37. [45]

    St3d: Self-training for unsupervised do- main adaptation on 3d object detection

    Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised do- main adaptation on 3d object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 3

    2021

  38. [46]

    St3d++: Denoised self-training for unsuper- vised domain adaptation on 3d object detection.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2022

    Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d++: Denoised self-training for unsuper- vised domain adaptation on 3d object detection.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2022. 3

    2022

  39. [47]

    3dssd: Point-based 3d single stage object detector.2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037–11045, 2020

    Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector.2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037–11045, 2020. 2

    2020

  40. [48]

    Center- based 3d object detection and tracking.CVPR, 2021

    Tianwei Yin, Xingyi Zhou, and Philipp Kr ¨ahenb¨uhl. Center- based 3d object detection and tracking.CVPR, 2021. 2, 4

    2021

  41. [49]

    Bi3d: Bi-domain active learn- ing for cross-domain 3d object detection

    Jiakang Yuan, Bo Zhang, Xiangchao Yan, Tao Chen, Botian Shi, Yikang Li, and Yu Qiao. Bi3d: Bi-domain active learn- ing for cross-domain 3d object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1497–1507, 2023. 3

    2023

  42. [50]

    Reg-tta3d: Better regres- sion makes better test-time adaptive 3d object detection

    Jiakang Yuan, Bo Zhang, Kaixiong Gong, Xiangyu Yue, Bo- tian Shi, Yu Qiao, and Tao Chen. Reg-tta3d: Better regres- sion makes better test-time adaptive 3d object detection. In ECCV, pages 197–213, 2024. 3, 6

    2024

  43. [51]

    Srdan: Scale- aware and range-aware domain adaptation network for cross- dataset 3d object detection

    Weichen Zhang, Wen Li, and Dong Xu. Srdan: Scale- aware and range-aware domain adaptation network for cross- dataset 3d object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6769–6779, 2021. 3

    2021

  44. [52]

    A survey of large language models, 2025

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu 8 Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong...

    2025

  45. [53]

    V oxelnet: End-to-end learning for point cloud based 3d object detection.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490–4499, 2017

    Yin Zhou and Oncel Tuzel. V oxelnet: End-to-end learning for point cloud based 3d object detection.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490–4499, 2017. 2

    2018

  46. [54]

    Zoox: Purpose-built autonomous mobility

    Zoox, Inc. Zoox: Purpose-built autonomous mobility. https://zoox.com/, 2025. Accessed: 2025-05-09. 1 9 Domain Adaptation for Different Sensor Configurations in 3D Object Detection Supplementary Material Figure 3. The total number of parameters for each layer

    2025

  47. [56]

    Recently, with the rapid development of large language models (LLMs) [17, 52], fine-tuning techniques themselves have also evolved

    Fine-tuning in Large Language Models Fine-tuning has traditionally been used in various tasks not only in academia but also in industry. Recently, with the rapid development of large language models (LLMs) [17, 52], fine-tuning techniques themselves have also evolved. After pr...

  48. [57]

    The Total Number of Parameters We first examined the total number of parameters in the model

    Additional analysis for fine-tuning 7.1. The Total Number of Parameters We first examined the total number of parameters in the model. The parameter count for each architectural module is as follows: • Encoder: 1,360 parameters • Backbone: 4,194,304 parameters • Neck: 599,552 ...

  49. [2025]

    Accessed: 2025-05-09. 1

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.