LiDARDustX: A LiDAR Dataset for Dusty Unstructured Road Environments

Boyang Zhao; Chenfeng Wei; Guotao Xie; Jiahua Xu; Qi Wu; Shenhong Wang; Si Zuo; Zeyu Yang

REVIEW 4 major objections 4 minor 22 references

LiDARDustX: A LiDAR Dataset for Dusty Unstructured Road Environments

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces a 30,000-frame, six-sensor LiDAR benchmark for dusty mining roads and shows that dust lowers 3D detection average precision by 9.9 to 18.1 points across six detectors.

desk verdict A genuinely useful dust-focused LiDAR dataset, but the paper's internal inconsistencies and the undefined dusty/dust-free split keep the headline dust-impact numbers from being fully established. read the letter →

arxiv 2505.21914 v1 pith:MIOTPN6E submitted 2025-05-28 cs.CV

Chenfeng Wei , Qi Wu , Si Zuo , Jiahua Xu , Boyang Zhao , Zeyu Yang , Guotao Xie , Shenhong Wang This is my paper

classification cs.CV

keywords LiDARdatasetdust-affectedscenes3Dobjectdetectionsemanticsegmentationunstructuredroadenvironmentsopen-pitmineautonomousdrivingbenchmarkmulti-taskperception

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

This paper tries to establish that dust is a distinct and measurable threat to LiDAR perception in unstructured driving, and it provides a public benchmark for studying that threat. The LiDARDustX dataset contains 30,000 LiDAR frames from six different sensors mounted on a mining truck, with 3D bounding boxes and pointwise semantic labels, and the paper reports that more than 80 percent of the frames are dust-affected. On a test set balanced between dusty and dust-free frames, six single-task detectors lose 9.9 to 18.1 points of average precision when dust is present, while segmentation degrades less. The authors argue that this benchmark fills a gap left by urban-focused driving datasets and enables development of perception methods that keep working when visibility is impaired by dust.

What carries the argument

The load-bearing object is the LiDARDustX dataset itself: 30,000 frames from six LiDAR sensor types, each carrying 7-degree-of-freedom 3D bounding boxes (center, dimensions, heading), tracking IDs, and 16-class pointwise semantic labels, split 70/30 into training and test sets with the test set balanced between dusty and dust-free frames. The annotation pipeline stitches the six sensors into one dense cloud, estimates local surface orientation to seed ground points, grows connected regions to separate ground from non-ground, clusters non-ground points against existing boxes, trains a segmentation model on these rough labels to produce pseudo-labels, and then has annotators refine the output into final labels. The benchmark experiments use this setup to compare six single-task 3D detectors, several segmentation models, and one multi-task baseline, making dust the manipulated variable.

What would settle it

Have an independent team manually label dust masks and bounding boxes for a random subset of, say, 500 test frames without seeing the dataset's labels, then recompute the dusty-versus-dust-free average-precision gap; if the gap changes materially or model rankings shift, the reported 9.9-to-18.1-point effect is an artifact of the annotation protocol rather than a property of dust.

Watch

Extended reading notes

Core claim

The central discovery is a quantified, controlled comparison: when the same detectors are evaluated on dusty versus dust-free frames from the same LiDAR sensors, every model loses performance, with average precision falling by 9.9 points for the most robust detector and 18.1 points for the least robust. Dust source also matters: naturally occurring dust mostly increases false positives and false negatives, while vehicle-stirred dust degrades heading-angle accuracy and inflates false positives. The paper further finds that a multi-task model that jointly learns segmentation and detection is among the most stable in dust, which it interprets as evidence that segmentation supervision helps when dust obscures object boundaries.

Load-bearing premise

The benchmark's conclusions rest on the assumption that the semi-automatic labeling process, in which a model trained on rough labels is corrected by human annotators, produces ground-truth dust and object labels accurate enough that the measured performance gaps are caused by dust, not by label noise or dust-label bias.

Editorial extensions

If this is right

Dust cannot be treated as ordinary sensor noise: all evaluated single-task detectors drop between 9.9 and 18.1 average-precision points on dusty frames, so dust-aware training or preprocessing is needed for mining-road autonomy.
The balanced dusty/dust-free test split gives future methods a single headline number, the AP drop, to optimize when claiming dust resistance.
Because segmentation holds up better than detection in dust, sharing segmentation features with detection, as the multi-task baseline does, is a workable route to more stable detection.
Dust-source type should be reported separately, since natural dust and vehicle-stirred dust produce different error patterns.

Reading between the lines

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

Extending the published benchmark from the three scored classes to the full 14 annotated classes would likely show larger dust penalties for small or rare objects, since the long-tailed class distribution already makes them harder to detect.
The six-sensor design could test whether dust tolerance tracks sensor line count or vertical field of view, but the paper does not report per-sensor accuracy, so that relationship remains an open question.
The error-cause statistics suggest targeted interventions: natural dust calls for denoising and hard-negative mining, while vehicle-stirred dust calls for heading-aware losses or orientation refinement.
The dataset's dust labels could also train a dust-density estimator as a gating module, letting a vehicle switch to a dust-mode detector when dust is present.

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.

Desk Editor's Note

read the letter

You should know this one for the dataset, not for the analysis around it. LiDARDustX is the first public mining-environment LiDAR dataset I know of with six sensor types, 30k annotated frames, 3D boxes and semantic labels including dust. That's a real gap and the resource itself looks worth having.

What's new: multi-sensor diversity and dust-specific semantic labels. The annotation pipeline (stitching, region growing, box-derived rough labels, model pseudo-labels, annotator correction) is a reasonable practical approach for a hard labeling problem. The benchmark covers six detectors and five segmenters, plus their LiSD multi-task model, and the reported dust-caused AP drop is plausible in direction.

But the paper is sloppier than it should be. Abstract says 80% dust frames, intro says 95%. Sequence count is 180 in Table I and Section III.E but 200 in Section III.A. Table IV contradicts its own text: text says natural dust drives FP/FN and vehicle dust drives orientation error; table has the opposite. No error bars or repeated runs anywhere. Most important, the dusty/dust-free split that produces the 9.9–18.1 AP drop has no stated criterion. If the split is derived from the dust semantic labels, and those labels come from a model trained on rough labels and then corrected, then errors in dust labels directly contaminate the independent variable. The paper needs to define the split, make the per-frame dust attribute public, and ideally show the split's sensitivity to label noise.

The self-citation issue is real but minor: LiSD is the authors' prior work and it wins, but the benchmark is new and the comparison is standard; that doesn't sink the paper. What sinks the dust-impact analysis as currently written is the missing partition definition, not the presence of their own model.

Who's this for? People working on perception for mining, off-road, or dusty environments; dataset papers are foundational for that niche. As a resource, it deserves referee time. The internal inconsistencies and split definition need a major revision before acceptance. I'd send it to review, but I'd want the authors to fix the numbers and release a dust attribute per frame.

Referee Report

4 major / 4 minor

Summary. The paper introduces LiDARDustX, a new LiDAR dataset for autonomous driving in dusty, unstructured environments such as open-pit mines and sand quarries. The dataset reportedly contains 30,000 annotated frames from six different LiDAR sensors, with 3D bounding boxes and point-cloud semantic segmentation labels, and is claimed to be predominantly composed of dust-affected scenes. The authors benchmark six 3D object detectors, five semantic segmentation networks, and one multi-task model, and analyze how dust degrades detection accuracy, reporting an average-precision drop between 9.9 and 18.1 points on dusty versus dust-free test subsets. The paper also presents an error-cause analysis based on a manual inspection of 100 error frames. The dataset is publicly announced via a GitHub link.

Significance. If the dataset is released as described and the numerical results are verified, LiDARDustX would fill a clear gap: currently available autonomous-driving datasets are overwhelmingly urban and structured, and dust is a realistic but underrepresented challenge for perception systems in mining and quarrying environments. The multi-sensor setup, the simultaneous support for detection and segmentation, the large annotated frame count, and the broad baseline evaluation are genuine strengths. The paper also makes a falsifiable claim about dust-induced performance degradation, which is valuable if the experimental protocol is made reproducible. The main weaknesses are internal statistical inconsistencies and an underspecified evaluation split, both of which affect the credibility of the headline quantitative claims.

major comments (4)

[Abstract, Section I, Section III.A, Section III.E, Table I] The dataset statistics are internally inconsistent. The abstract states that 'over 80%' of the dataset comprises dust-affected scenes, while the introduction's contribution bullet states 'over 95%'. Section III.A says that 200 dust-rich sequences were selected, but Table I and Section III.E both report 180 sequences. Additionally, 180 sequences of approximately 20 seconds at the stated 5 Hz keyframe rate would yield roughly 18,000 frames, not the claimed 30,000. These discrepancies affect the dataset's identity and must be resolved with precise definitions and a single set of consistent numbers.
[Section III.E and Section IV.D, Table III] The central result that dust degrades detection AP by 9.9–18.1 points depends on partitioning the test set into 'dusty' and 'dust-free' subsets, but the paper never defines how this partition is made. Section III.E only says the test set is 'comprised of an equal amount of dusty and dust-free data', and Section IV.D says performance was 'assessed on dust-affected and dust-free subsets'; there is no frame-level dust attribute, no threshold, and no manual protocol. The only dust signal in the annotation pipeline is the semantic Dust class, which is produced by a segmentation model trained on rough labels and then corrected by annotators (Section III.C.4–6). If the split inherits errors from these pseudo-labels, the independent variable of Table III is contaminated. Please specify the partition criterion, release per-frame dust labels, and report dust-point fraction statistics for both subsets.
[Section V and Table IV] The error-cause analysis is internally contradictory. The text states that natural dust mainly increases false positives and false negatives without significantly affecting orientation accuracy, while vehicle-induced dust degrades the precision of orientation and increases false positives. Table IV shows the opposite pattern: vehicular activities have FP=21, FN=25, and orientation error=0, whereas naturally occurring dust has FP=19, FN=2, and orientation error=43. The table and the text cannot both be correct. The sampling protocol for the 'subjective examination' (100 frames, no seed, no detailed error taxonomy) is also not described. Please correct the discrepancy and provide a reproducible protocol for the error analysis.
[Section III.C and Table III] The paper asserts that the annotations are high-precision and uses Table III to support quantitative conclusions, but it provides no annotation-quality metrics and no run-to-run variability estimates. There are no inter-annotator agreement statistics, no validation protocol for the pseudo-label refinement step, and no error bars or multiple-seed results for the benchmark numbers. Without such evidence, the observed AP drops cannot be separated from label noise and optimization variance. Please add quantitative annotation-quality measures and report standard deviations or confidence intervals for the main benchmark results.

minor comments (4)

[References] Reference [12] is cited for CenterPoint, but the given title, 'Centerpoints: A link between optimization and convex geometry', is not the CenterPoint 3D object detection paper. Please cite the correct reference.
[Table II and Section IV.C] The name of the multi-task model is written as 'LisD' in Table II and as 'LiSD' in the text; please standardize the spelling.
[Section III.D] The paper mentions a 1:30 ratio between the least and most annotated classes, but the underlying counts are not stated numerically. Please provide the exact point counts for the smallest and largest semantic classes so the ratio can be verified.
[Dataset release] The dataset is announced through a GitHub link, but the paper does not specify a versioned release, a persistent identifier such as a DOI, or a license. These details should be added for reproducibility and community use.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: LiDARDustX is a new dataset with empirical benchmarks; the only self-reference is the authors' LiSD baseline, which is evaluated on the new data, not assumed.

full rationale

This paper contains no derivation chain whose output is equivalent to its inputs. The contributions are a new 30,000-frame LiDAR dataset, annotations, and benchmark evaluations. The dusty/dust-free test split is not fully specified (Sec. III.E only says the test set is 'comprised of an equal amount of dusty and dust-free data'), and the dust labels originate from a model-trained pseudo-label pipeline corrected by annotators (Sec. III.C.5-6); that is a measurement-validity concern about the independent variable, but the paper does not define the split in terms of the benchmark outputs, so no equation or fitted parameter is renamed as a prediction. The strongest self-referential element is that the top-performing multi-task model LiSD is the authors' prior work [20], and the paper highlights its superiority; however, LiSD's numbers are measured on the new dataset rather than imported from the citation, so the citation is not load-bearing. The internal inconsistency between the Table IV error-cause statistics and the prose description of natural versus vehicle dust is a correctness issue, not circularity. No self-definitional step, fitted-input-as-prediction, or uniqueness-imported-by-authors pattern is present.

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

The dataset is empirical; no new physical parameters or entities are introduced. The assumptions are about data quality and annotation reliability, which are the paper's true load-bearing premises.

assumptions (4)

domain assumption The six LiDAR sensors are synchronized and extrinsically calibrated so the stitched point clouds used for annotation are spatially consistent.
Section III.C.1 relies on stitching for annotation.
domain assumption The semantic class 'dust' can be consistently distinguished from obstacles and ground by the annotators and the pseudo-label model.
Section III.C.6 and Section V interpret dust as a cleanly separable class.
domain assumption The 70/30 training/test split preserves the dusty and dust-free frame distribution so the dust impact comparison in Section IV.D is fair.
Section III.E describes the split; Section IV.D compares dusty vs dust-free subsets without reporting per-subset sizes.
standard math Standard detection and segmentation metrics (mAP at 0.7/0.5 IoU, mIoU) are computed correctly on the released test set.
Section III.E defines the metrics; no code is provided to verify them.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LiDARDustX: A LiDAR Dataset for Dusty Unstructured Road Environments." pith.science (2026). https://pith.science/paper/MIOTPN6E

@misc{pith2026250521914,
  author       = {Pith},
  title        = {Pith review of: LiDARDustX: A LiDAR Dataset for Dusty Unstructured Road Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MIOTPN6E}},
  note         = {Machine review of arXiv:2505.21914}
}

read the original abstract

Autonomous driving datasets are essential for validating the progress of intelligent vehicle algorithms, which include localization, perception, and prediction. However, existing datasets are predominantly focused on structured urban environments, which limits the exploration of unstructured and specialized scenarios, particularly those characterized by significant dust levels. This paper introduces the LiDARDustX dataset, which is specifically designed for perception tasks under high-dust conditions, such as those encountered in mining areas. The LiDARDustX dataset consists of 30,000 LiDAR frames captured by six different LiDAR sensors, each accompanied by 3D bounding box annotations and point cloud semantic segmentation. Notably, over 80% of the dataset comprises dust-affected scenes. By utilizing this dataset, we have established a benchmark for evaluating the performance of state-of-the-art 3D detection and segmentation algorithms. Additionally, we have analyzed the impact of dust on perception accuracy and delved into the causes of these effects. The data and further information can be accessed at: https://github.com/vincentweikey/LiDARDustX.

Figures

Figures reproduced from arXiv: 2505.21914 by the authors.

**Figure 2.** Collection Platform LiDAR Position Schematic. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The LiDARDustX dataset showcases point clouds from various LiDAR sensors. The points in the cycle are zoomed [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 6.** Figure 6: The number of points in different semantic categories. [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 4.** Figure 4: The methodologies for point cloud segmentation [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: The number of 3D box annotations in different [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

[1]

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

A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in2012 IEEE conference on computer vision and pattern recognition, pp. 3354–3361, IEEE, 2012

2012
[2]

nuscenes: A multimodal dataset for autonomous driving,

H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11621–11631, 2020

2020
[3]

Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,

S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. R. Qi, Y . Zhou,et al., “Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9710–9719, 2021

work page 2021
[4]

Argoverse: 3d tracking and forecasting with rich maps,

M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan,et al., “Argoverse: 3d tracking and forecasting with rich maps,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 8748–8757, 2019

work page 2019
[5]

A* 3d dataset: Towards autonomous driving in challenging environments,

Q.-H. Pham, P. Sevestre, R. S. Pahwa, H. Zhan, C. H. Pang, Y . Chen, A. Mustafa, V . Chandrasekhar, and J. Lin, “A* 3d dataset: Towards autonomous driving in challenging environments,” in2020 IEEE In- ternational conference on Robotics and Automation (ICRA), pp. 2267– 2273, IEEE, 2020

work page 2020
[6]

A2d2: Audi autonomous driving dataset,

J. Geyer, Y . Kassahun, M. Mahmudi, X. Ricou, R. Durgesh, A. S. Chung, L. Hauswald, V . H. Pham, M. Mühlegg, S. Dorn,et al., “A2d2: Audi autonomous driving dataset,”arXiv preprint arXiv:2004.06320, 2020

arXiv 2004
[7]

One million scenes for autonomous driving: Once dataset,

J. Mao, M. Niu, C. Jiang, H. Liang, J. Chen, X. Liang, Y . Li, C. Ye, W. Zhang, Z. Li,et al., “One million scenes for autonomous driving: Once dataset,”arXiv preprint arXiv:2106.11037, 2021

arXiv 2021
[8]

Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,

J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 9297–9307, 2019

work page 2019

Show all 22 references

[9]

Automine: An unmanned mine dataset,

Y . Li, Z. Li, S. Teng, Y . Zhang, Y . Zhou, Y . Zhu, D. Cao, B. Tian, Y . Ai, Z. Xuanyuan,et al., “Automine: An unmanned mine dataset,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21308–21317, 2022

2022
[10]

Pointpillars: Fast encoders for object detection from point clouds,

A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2019

2019
[11]

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

S. Shi, C. Guo, L. Jiang, Z. Wang, J. Shi, X. Wang, and H. Li, “Pv-rcnn: Point-voxel feature set abstraction for 3d object detection,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020

2020
[12]

Centerpoints: A link between optimization and convex geometry,

A. Basu and T. Oertel, “Centerpoints: A link between optimization and convex geometry,”arXiv: Optimization and Control,arXiv: Opti- mization and Control, Nov 2015

2015
[13]

Pillarnet: Real-time and high-performance pillar-based 3d object detection,

G. Shi, R. Li, and C. Ma, “Pillarnet: Real-time and high-performance pillar-based 3d object detection,” inEuropean Conference on Com- puter Vision, pp. 35–52, Springer, 2022

2022
[14]

V oxelnext: Fully sparse voxelnet for 3d object detection and tracking,

Y . Chen, J. Liu, X. Zhang, X. Qi, and J. Jia, “V oxelnext: Fully sparse voxelnet for 3d object detection and tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21674–21683, 2023

2023
[15]

Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,

X. Bai, Z. Hu, X. Zhu, Q. Huang, Y . Chen, H. Fu, and C.-L. Tai, “Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1090–1099, 2022

2022
[16]

Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

C. Qi, L. Yi, H. Su, and L. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,”Cornell University - arXiv,Cornell University - arXiv, Jun 2017

2017
[17]

Randla-net: Efficient semantic segmentation of large-scale point clouds,

Q. Hu, B. Yang, L. Xie, S. Rosa, Y . Guo, Z. Wang, N. Trigoni, and A. Markham, “Randla-net: Efficient semantic segmentation of large-scale point clouds,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020

2020
[18]

Cenet: Consolidation-and-exploration network for continuous domain adapta- tion,

C. Zhang, Y . Cheng, P. Wei, H. He, and J. Chen, “Cenet: Consolidation-and-exploration network for continuous domain adapta- tion,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Jun 2022

2022
[19]

Cylin- der3d: An effective 3d framework for driving-scene lidar semantic seg- mentation.,

H. Zhou, X. Zhu, X. Song, Y . Ma, Z. Wang, H. Li, and D. Lin, “Cylin- der3d: An effective 3d framework for driving-scene lidar semantic seg- mentation.,”arXiv: Computer Vision and Pattern Recognition,arXiv: Computer Vision and Pattern Recognition, Aug 2020

2020
[20]

Lisd: An efficient multi- task learning framework for lidar segmentation and detection,

J. Xu, S. Zuo, C. Wei, and W. Zhou, “Lisd: An efficient multi- task learning framework for lidar segmentation and detection,”arXiv preprint arXiv:2406.07023, 2024

2024 arXiv
[21]

The pascal visual object classes (voc) challenge,

M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisser- man, “The pascal visual object classes (voc) challenge,”International journal of computer vision, vol. 88, pp. 303–338, 2010

2010
[22]

V oxelnet: End-to-end learning for point cloud based 3d object detection,

Y . Zhou and O. Tuzel, “V oxelnet: End-to-end learning for point cloud based 3d object detection,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

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

A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in2012 IEEE conference on computer vision and pattern recognition, pp. 3354–3361, IEEE, 2012

2012

[2] [2]

nuscenes: A multimodal dataset for autonomous driving,

H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11621–11631, 2020

2020

[3] [3]

Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,

S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. R. Qi, Y . Zhou,et al., “Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9710–9719, 2021

work page 2021

[4] [4]

Argoverse: 3d tracking and forecasting with rich maps,

M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan,et al., “Argoverse: 3d tracking and forecasting with rich maps,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 8748–8757, 2019

work page 2019

[5] [5]

A* 3d dataset: Towards autonomous driving in challenging environments,

Q.-H. Pham, P. Sevestre, R. S. Pahwa, H. Zhan, C. H. Pang, Y . Chen, A. Mustafa, V . Chandrasekhar, and J. Lin, “A* 3d dataset: Towards autonomous driving in challenging environments,” in2020 IEEE In- ternational conference on Robotics and Automation (ICRA), pp. 2267– 2273, IEEE, 2020

work page 2020

[6] [6]

A2d2: Audi autonomous driving dataset,

J. Geyer, Y . Kassahun, M. Mahmudi, X. Ricou, R. Durgesh, A. S. Chung, L. Hauswald, V . H. Pham, M. Mühlegg, S. Dorn,et al., “A2d2: Audi autonomous driving dataset,”arXiv preprint arXiv:2004.06320, 2020

arXiv 2004

[7] [7]

One million scenes for autonomous driving: Once dataset,

J. Mao, M. Niu, C. Jiang, H. Liang, J. Chen, X. Liang, Y . Li, C. Ye, W. Zhang, Z. Li,et al., “One million scenes for autonomous driving: Once dataset,”arXiv preprint arXiv:2106.11037, 2021

arXiv 2021

[8] [8]

Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,

J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 9297–9307, 2019

work page 2019