Pith. sign in

REVIEW 5 major objections 5 minor 25 references

DINOSTAR: Deep Iterative Neural Object Detector Self-Supervised Training for Roadside LiDAR Applications

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that deep object detectors for roadside LiDAR point clouds can be trained entirely without human labels, using noisy labels from statistical teacher models and an iterative student-teacher loop, and that this reaches…

desk verdict The self-supervised teacher-student pipeline for roadside LiDAR is plausible and the multi-teacher data combination helps, but the paper's headline claim of parity with human-supervised training is unsupported because no human-supervised baseline is ever trained. read the letter →

arxiv 2501.17076 v1 pith:OS4WCLTD submitted 2025-01-28 cs.CV

Muhammad Shahbaz , Shaurya Agarwal This is my paper · ORCID
classification cs.CV
keywords self-supervisedlearningobjectdetectionroadsideLiDARpointcloudstudent-teachermodelingbackgroundfilteringDBSCANpedestrian
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

Roadside LiDAR sensors produce raw point clouds that are costly to label by hand, and the paper tries to remove that bottleneck. It proposes a teacher–student pipeline in which statistical teacher models—built from a distance-histogram background filter, density-based clustering, bounding-box fitting, and a shape heuristic—generate noisy object labels for unannotated frames, and an off-the-shelf deep point-cloud detector is trained as a student on the union of those labels. The central empirical claim is that this self-supervised route reaches performance comparable to human-annotated training: on the Pedsafe test set, pedestrian detection mAP@0.25 rises from 40.77 when the student trains on Pedsafe teacher labels alone to 61.37 when teacher labels from the APLS campus dataset are added. If that holds, roadside perception systems can be scaled to new locations and perspectives without spending the time and money that human labeling requires.

What carries the argument

The load-bearing mechanism is the teacher-student loop, with the D-HistDPP background filter as the key new component. D-HistDPP computes, for each of Ntotal zero-padded point positions, a histogram of distances to that position across Nquery query frames; the Ntall tallest histogram bins are taken to be the background locations hit by that ray, and any point whose distance differs from such a bin by less than Dthreshold is removed as background. Cleaned foreground points feed DBSCAN clustering, clusters are enclosed in minimal bounding boxes, and a shape heuristic (vehicle base-length larger than height; pedestrian height larger than base-length) assigns the class. The student is a sparse-convolutional point-cloud detector trained to regress those boxes and classes from the union of labels produced by multiple teachers, and iterative self-training recycles the student's own predictions as ground truth for a successor model.

What would settle it

Train the same sparse-convolutional detector on the same Pedsafe training frames with human-annotated boxes instead of teacher-generated boxes, and evaluate on the same 1000-frame test split. If that human-supervised baseline's mAP@0.25 is substantially above 61.37, the claim of comparable performance without human annotations fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a deep object detector can be trained for roadside LiDAR without any human-annotated ground truth and still perform comparably to supervised training. The teacher side implements a traditional four-step detector: D-HistDPP removes background by building a per-point distance histogram over the first Nquery frames, retaining the Ntall tallest bins as background locations and discarding points within Dthreshold; the remaining foreground is clustered with DBSCAN; clusters are wrapped in minimal bounding boxes; and a simple shape rule labels vehicles when base-length exceeds height and pedestrians otherwise. A sparse-convolutional point-cloud detector (the student) is then trained on these noisy labels, and training can be repeated iteratively by using the student's own predictions as labels for its successor. The evidence the paper highlights is the pedestrian-detection gain on Pedsafe: 40.77 mAP@0.25 with one teacher's labels versus 61.37 mAP@0.25 when teacher labels from APLS are added, which the authors present as comparable to detectors trained on human annotations.

Load-bearing premise

The background filter assumes that for every zero-padded point position, the distances to static background objects across the query frames pile up in the tallest histogram bins, so that removing points within a fixed threshold of those bins leaves foreground objects intact; if foliage moves, the sensor jitters more than the threshold, or zero-padding misaligns points, foreground points get erased or background noise survives.

Editorial extensions

If this is right

  • Pedestrian and vehicle detection for roadside LiDAR can be trained from raw, unlabeled point-cloud streams, removing the human-labeling bottleneck that currently limits scaling to many sensor placements.
  • Adding auto-annotated frames from a second, differently placed sensor (APLS) to a first (Pedsafe) improves pedestrian mAP@0.25 by about 20 points, consistent with the paper's claim that an ensemble of teachers diversifies point-cloud representations.
  • Because the student is an off-the-shelf detector, the labeling pipeline can remain fixed while the underlying detector architecture improves.
  • Iterative self-training gives diminishing returns after roughly three generations, suggesting a natural stopping point for the student-teacher loop.
  • Cross-dataset tests on A9 and IPS300+ show that very different sensor perspectives and scene types still cause large performance drops, so the method's transfer across radically new installations remains limited.

Reading between the lines

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

  • The large gain from adding a second dataset suggests a scaling law worth testing: how much pedestrian mAP improves as the number of independently placed teachers increases, and at what point teacher noise makes additional data unhelpful.
  • The framework's dependence on the D-HistDPP assumption that background distances concentrate in tall bins implies it will struggle on moving or shaking sensors; a testable fix is to replace the fixed Dthreshold with a per-bin adaptive tolerance estimated from the histogram spread.
  • The paper does not compare against a human-annotated baseline on the same frames nor against plain student self-training, so a reader could test whether the teacher labels add value beyond pseudo-labeling by the student itself.
  • The shape heuristic only covers vehicles and pedestrians; extending it to cyclists or trucks would be a direct test of whether the heuristic labeling bottleneck, rather than the detector, limits the class set.
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

5 major / 5 minor

Summary. The manuscript proposes DINOSTAR, a teacher–student framework for training point-cloud object detectors without human-annotated training labels. A hand-designed teacher pipeline (pre-processing, D-HistDPP background filtering, DBSCAN clustering, heuristic bounding-box fitting, and heuristic classification) generates noisy labels; a SECOND detector is then trained on combined teacher labels from the authors' Pedsafe dataset and the public APLS dataset, and evaluated on human-labeled test frames from Pedsafe, A9, and IPS300+. The main positive result is pedestrian mAP@0.25 improving from 40.77 to 61.37 when APLS teacher labels are added. The central claim of parity with human-supervised training is not tested, because no human-supervised baseline is trained.

Significance. The problem is relevant and the idea of ensembling statistical, self-supervised teachers to train a student detector is reasonable. The use of externally human-labeled test frames gives independent grounding to the reported numbers. If the parity claim were supported, the framework would be a useful way to scale roadside LiDAR object detection without annotation cost. However, the absence of a supervised baseline, the missing iterative-training evidence, the ill-defined background-filter pseudocode, and the near-zero cross-dataset results prevent me from treating the main claims as established.

major comments (5)
  1. [Abstract and Table III] The central claim that the framework 'achieves comparable performance to deep object detectors trained on human-annotated labels' is unsupported because no human-supervised baseline is trained. Table III compares only students trained on teacher-generated labels: Pedsafe-only versus Pedsafe plus APLS. The improvement from 40.77 to 61.37 mAP@0.25 shows that adding more teacher-annotated data helps, but it does not quantify any gap to human supervision. Please train a SECOND model on the same training frames with human labels (or a held-out annotated subset) and report the same metrics; if no human labels are available for training, state that explicitly and revise the abstract's parity claim.
  2. [III-A2, Algorithm 1 and Algorithm 2] The background-filter pseudocode is internally inconsistent. Algorithm 1 accumulates summed distances into Hdistance (line 15) and then line 18 computes 'mean(Hdistance_i,k)', which does not produce a meaningful bin height because the number of points per bin is not tracked. Algorithm 2 then sorts these accumulated sums as 'heights' and uses them as representative distances for background subtraction. As written, the algorithm is ill-defined. Please correct the pseudocode to store per-bin counts or a well-defined distance representative, and update Algorithm 2 accordingly. Additionally, because frames are zero-padded to Ntotal points, the padded positions have zero distance; clarify how the filter avoids treating those zero-distance entries as background and erasing real foreground points.
  3. [IV-A and Section V] The iterative component that gives the method its name is not evaluated. Section V states that the iterative process improves accuracy initially and that after three iterations accuracy no longer improves significantly, but no table or figure reports mAP for successive iterations. Please present per-iteration results on the same human-labeled test set, including iteration 0 (pre-iteration), 1, 2, and 3, so the reader can verify that iterative self-training actually provides the claimed benefit.
  4. [Table III, A9 and IPS300+ rows] The cross-dataset results are effectively zero and are not discussed in the text. On A9, all mAP values are 0.00 for both training configurations, and the two Recall@0.3 values are identical (0.861); on IPS300+, mAP@0.25 is at most 6.28. These results contradict the broad 'roadside LiDAR applications' framing in the abstract and point to strong location/perspective specificity. Please either narrow the stated scope to same-sensor/same-location deployment or provide domain-adaptation evidence. Also clarify why the A9 Recall@0.3 values are identical across training configurations.
  5. [Section IV and Section III-A] All reported results appear to come from a single run with no error bars, despite the large number of teacher hyperparameters (Nquery, Nbin, Ntall, Dthreshold, DBSCAN epsilon, DBSCAN MinPts, lmin, hmin, beta_min, and the data-unification scaling/centering factors). Without a sensitivity analysis or multiple seeds, it is difficult to know whether the key 61.37 mAP@0.25 result is robust or a result of parameter fitting. Please report at least mean and standard deviation over multiple runs and an ablation over the main teacher hyperparameters.
minor comments (5)
  1. [Throughout] The manuscript contains unfinished placeholders that must be resolved before any review: '[Link to our Github Repo]', 'refer to OpenPCDet paper', '[reference]', and '[mention additional models]'.
  2. [Table I] Several table entries are incomplete or uncertain, such as the '6ft?' height and '10Hz?' frequency for APLS; please complete all specifications or mark unavailable values clearly.
  3. [Section III-B] The iterative training process is described only at a high level; please provide the training hyperparameters for the student model in each iteration, including learning rate schedule, number of epochs, and how the student-generated labels are filtered or confidence-thresholded.
  4. [Contributions, Section I] Contribution 4 states 'It is first to evaluates the results on IPS300+'; this should be rewritten grammatically and the novelty claim checked against existing IPS300+ benchmarks, since prior work on IPS300+ may already report detection results.
  5. [References] References [23] and [24] appear to be duplicate entries of the same paper by Zhou et al. with different venues; please consolidate and verify all citations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the teacher-generated training labels and the human-labeled test targets are independent inputs, so the reported improvements are not forced by construction.

full rationale

The paper's derivation chain is self-contained in the sense required for circularity analysis. Raw point-cloud frames are processed by statistically modeled teachers that use background filtering (D-HistDPP), DBSCAN clustering, bounding-box fitting, and heuristic classification to produce noisy labels. The student detector (SECOND) is trained on those teacher labels, and its performance is then evaluated on held-out frames with human-annotated ground truth (Pedsafe test frames, A9, IPS300+). The central quantitative claims—mAP and recall on pedestrian detection—are therefore measured against external human annotations, not against the teacher labels or the student's own outputs. No fitted parameter is renamed as a prediction; the improvement from adding APLS teacher labels is an empirical data effect rather than an identity. No load-bearing self-citation appears: the only author-specific resource is the Pedsafe dataset, which is human-labeled and used as an external benchmark. The iterative self-training idea is described but no circular result is claimed from it, and the paper notes that gains saturate after three iterations. The abstract's assertion of 'comparable performance to deep object detectors trained on human-annotated labels' is not actually demonstrated by a same-architecture supervised baseline, and several references and the GitHub link are placeholders; these are evidentiary and completeness gaps, not circularity. The D-HistDPP background filter relies on a stationarity assumption that could hurt performance on difficult scenes, but that is a robustness concern, not a definitional loop. Overall, the evaluation protocol breaks any direct circular feedback from the teacher-student pipeline into the final numbers.

Assumptions & free parameters 11 free parameters · 5 assumptions · 0 invented entities

The central pipeline relies on many hand-tuned hyperparameters that are not reported, all of which affect label quality. No new physical entities are introduced. The domain assumptions about static scenes, object clustering, and size-based classification are load-bearing and untested outside the authors' own dataset.

free parameters (11)
  • Ntotal (zero padding size)
    Fixed number of points per frame in pre-processing (Section III-A1). Value not reported.
  • Nquery (number of query frames)
    Used to build the distance histogram (Equation 2, Section III-A2). Value not reported.
  • Nbin (number of histogram bins)
    Parameter of D-HistDPP in Algorithm 1. Value not reported.
  • Ntall (number of tallest bins)
    Number of background locations per point-ray in Algorithm 2. Value not reported.
  • Dthreshold (background distance threshold)
    Radius around background locations for subtraction in Algorithm 2. Value not reported.
  • DBSCAN epsilon
    Distance threshold for DBSCAN clustering (Section III-A3). Value not reported.
  • DBSCAN MinPts
    Minimum points to form a cluster (Section III-A3). Value not reported.
  • lmin (minimum bbox base length)
    Bounding-box validation heuristic (Section III-A4). Value not reported.
  • hmin (minimum bbox height)
    Bounding-box validation heuristic (Section III-A4). Value not reported.
  • beta_min (minimum base-height difference)
    Bounding-box validation heuristic (Section III-A4). Value not reported.
  • Data unification scaling and centering factors
    Used to combine Pedsafe and APLS in Section IV. Values not reported.
assumptions (5)
  • domain assumption Static background can be modeled by a per-ray distance histogram over Nquery frames, and the Ntall tallest bins capture background distances.
    Introduced in Section III-A2. If background moves or sensor jitter exceeds Dthreshold, foreground estimation fails.
  • domain assumption DBSCAN clusters of foreground points correspond to whole objects of interest.
    Section III-A3. Assumes one cluster per object and that noisy points are handled by the chosen parameters.
  • domain assumption Vehicles are wider than tall, pedestrians are taller than wide, and these size heuristics are sufficient for classification.
    Section III-A4. Classification errors in the teacher propagate directly to the student's training labels.
  • domain assumption Different roadside LiDAR datasets can be unified by centering and scaling so that objects have roughly the same size.
    Section IV. Used to combine Pedsafe and APLS. Zero mAP on A9 and IPS300+ suggests this assumption fails across perspectives.
  • standard math DBSCAN and point-cloud distance computations follow standard definitions.
    Used in Algorithm 1 and clustering; no new mathematics is introduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DINOSTAR: Deep Iterative Neural Object Detector Self-Supervised Training for Roadside LiDAR Applications." pith.science (2026). https://pith.science/paper/OS4WCLTD

@misc{pith2026250117076,
  author       = {Pith},
  title        = {Pith review of: DINOSTAR: Deep Iterative Neural Object Detector Self-Supervised Training for Roadside LiDAR Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OS4WCLTD}},
  note         = {Machine review of arXiv:2501.17076}
}
read the original abstract

Recent advancements in deep-learning methods for object detection in point-cloud data have enabled numerous roadside applications, fostering improvements in transportation safety and management. However, the intricate nature of point-cloud data poses significant challenges for human-supervised labeling, resulting in substantial expenditures of time and capital. This paper addresses the issue by developing an end-to-end, scalable, and self-supervised framework for training deep object detectors tailored for roadside point-cloud data. The proposed framework leverages self-supervised, statistically modeled teachers to train off-the-shelf deep object detectors, thus circumventing the need for human supervision. The teacher models follow fine-tuned set standard practices of background filtering, object clustering, bounding-box fitting, and classification to generate noisy labels. It is presented that by training the student model over the combined noisy annotations from multitude of teachers enhances its capacity to discern background/foreground more effectively and forces it to learn diverse point-cloud-representations for object categories of interest. The evaluations, involving publicly available roadside datasets and state-of-art deep object detectors, demonstrate that the proposed framework achieves comparable performance to deep object detectors trained on human-annotated labels, despite not utilizing such human-annotations in its training process.

Figures

Figures reproduced from arXiv: 2501.17076 by the authors.

Figure 1
Figure 1. We gather point cloud data from multiple location [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Label Generation Pipeline: The teacher model (left) uses a novel background filtering method followed by traditional [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Iterative Training for Improving Student Belief [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages

  1. [1]

    Bai, S.P

    Z. Bai, S.P. Nayak, X. Zhao, G. Wu, M.J. Barth, X. Qi, and K. Oguchi. Cyber mobility mirror: Deep learning-based real-time 3d object perception and reconstruction using roadside lidar. arXiv preprint arXiv:2202.13505, 2022

    arXiv 2022

  2. [2]

    Apollo: A dataset profiling and operator modeling system

    Tasos Bakogiannis, Ioannis Giannakopoulos, Dimitrios Tsoumakos, and Nectarios Koziris. Apollo: A dataset profiling and operator modeling system. In Proceedings of the 2019 International Conference on Management of Data , pages 1869–1872, 2019

    work page 2019

  3. [3]

    Unsupervised lidar-based 3d object detection using infrastructure sensors, 2022

    Marcel Brucker. Unsupervised lidar-based 3d object detection using infrastructure sensors, 2022

    work page 2022

  4. [4]

    Lumpi: The leibniz university multi-perspective intersection dataset

    Steffen Busch, Christian Koetsier, Jeldrik Axmann, and Claus Brenner. Lumpi: The leibniz university multi-perspective intersection dataset. In 2022 IEEE Intelligent Vehicles Symposium (IV) , pages 1127–1134. IEEE, 2022

    work page 2022

  5. [5]

    nuscenes: A multimodal dataset for autonomous driving

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

    2020

  6. [6]

    Big data deep learning: challenges and perspectives

    Xue-Wen Chen and Xiaotong Lin. Big data deep learning: challenges and perspectives. IEEE access, 2:514–525, 2014

    work page 2014

  7. [7]

    A9-dataset: Multi-sensor infrastructure-based dataset for mobility research

    Christian Creß, Walter Zimmer, Leah Strand, Maximilian Fortkord, Siyi Dai, Venkatnarayanan Lakshminarasimhan, and Alois Knoll. A9-dataset: Multi-sensor infrastructure-based dataset for mobility research. In 2022 IEEE Intelligent Vehicles Symposium (IV) , pages 965–970. IEEE, 2022

    work page 2022

  8. [8]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    Martin Ester, Hans-Peter Kriegel, J ¨org Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, volume 96, pages 226–231, 1996

    1996

Show all 25 references
  1. [9]

    Vision meets robotics: The kitti dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231–1237, 2013

    2013

  2. [10]

    Deep learning for 3d point clouds: A survey

    Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mo- hammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338– 4364, 2020

    2020

  3. [11]

    Revisiting unreasonable effectiveness of data in deep learning era

    Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision , pages 843–852, 2017

    2017

  4. [12]

    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, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vi...

    2020

  5. [13]

    Ob- ject detection based on roadside lidar for cooperative driving automation: a review

    Pengpeng Sun, Chenghao Sun, Runmin Wang, and Xiangmo Zhao. Ob- ject detection based on roadside lidar for cooperative driving automation: a review. Sensors, 22(23):9316, 2022

    2022

  6. [14]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521–1528. IEEE, 2011

    2011

  7. [15]

    G. Wang, J. Wu, T. Xu, and B. Tian. 3d vehicle detection with rsu lidar for autonomous mine. IEEE Transactions on Vehicular Technology, 70:344–355, 2021

    2021

  8. [16]

    Ips300+: a challenging multi-modal data sets for intersection perception system

    Huanan Wang, Xinyu Zhang, Zhiwei Li, Jun Li, Kun Wang, Zhu Lei, and Ren Haibing. Ips300+: a challenging multi-modal data sets for intersection perception system. In 2022 International Conference on Robotics and Automation (ICRA) , pages 2539–2545. IEEE, 2022

    2022

  9. [17]

    Automatic background filtering method for roadside lidar data

    Jianqing Wu, Hao Xu, Yuan Sun, Jianying Zheng, and Rui Yue. Automatic background filtering method for roadside lidar data. Trans- portation Research Record, 2672(45):106–114, 2018

    2018

  10. [18]

    Pandaset: Advanced sensor suite dataset for autonomous driving

    Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), page...

    2021

  11. [19]

    Rope3d: the roadside perception dataset for autonomous driving and monocular 3d object detection task

    Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: the roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2022

  12. [20]

    Baai-vanjee roadside dataset: Towards the connected automated vehicle highway technologies in challenging environments of china

    Deng Yongqiang, Wang Dengjiang, Cao Gang, Ma Bing, Guan Xijia, Wang Yajun, Liu Jianchao, Fang Yanming, and Li Juanjuan. Baai-vanjee roadside dataset: Towards the connected automated vehicle highway technologies in challenging environments of china. arXiv preprint arXiv:2105.14...

    2021 arXiv

  13. [21]

    Dair-v2x: A large- scale dataset for vehicle-infrastructure cooperative 3d object detection

    Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large- scale dataset for vehicle-infrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and...

    2022

  14. [22]

    Zhang, W

    J. Zhang, W. Xiao, and J.P. Mills. Optimizing moving object trajectories from roadside lidar data by joint detection and tracking. Remote Sensing, 14:2124, 2022

    2022

  15. [23]

    S. Zhou, H. Xu, G. Zhang, T. Ma, and Y . Yang. Leveraging deep convolutional neural networks pre-trained on autonomous driving data for vehicle detection from roadside lidar data. IEEE Trans. Intell. Transp. Syst., 23:22367–22377, 2022

    2022

  16. [24]

    S. Zhou, H. Xu, G. Zhang, T. Ma, and Y . Yang. Leveraging deep convolutional neural networks pre-trained on autonomous driving data for vehicle detection from roadside lidar data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1052...

    2022

  17. [25]

    Zimmer, M

    W. Zimmer, M. Grabler, and A. Knoll. Real-time and robust 3d object detection within road-side lidars using domain adaptation. arXiv preprint arXiv:2204.00132, 2022

    2022 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.