Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Safety-Critical Learning for Long-Tail Events: The TUM Traffic Accident Dataset

T0 review · 4 major / 5 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 The TUM Traffic Accident dataset supplies 48,144 labeled frames of real highway crashes and a hybrid detector, Accid3nD, that flags accidents in real time.

desk verdict Genuinely new real-world highway accident dataset with dense 3D annotations, but the method evaluation is missing—claimed SOTA and ablations never appear; still worth a serious peer review. read the letter →

arxiv 2508.14567 v1 pith:YV2VGPHN submitted 2025-08-20 cs.CV

Walter Zimmer , Ross Greer , Xingcheng Zhou , Rui Song , Marc Pavel , Daniel Lehmberg , Ahmed Ghita , Akshay Gopalkrishnan
show 2 more authors
Mohan Trivedi Alois Knoll
This is my paper · ORCID
classification cs.CV
keywords TUMTrafficAccidentdatasetroadsideperceptiondetectionlong-tailsafetyevents3Dobjectannotationmulti-modalsensorfusiontrajectory-basedreal-worldhighwaycrashes
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 is trying to establish that naturally occurring highway accidents can be captured, densely labeled, and used to train and evaluate accident detection. The TUM Traffic Accident dataset provides 48,144 frames of real crashes from four roadside cameras and LiDARs, with 294,924 2D and 93,012 3D boxes plus track IDs in ten sequences. The authors propose Accid3nD, which first applies trajectory thresholds to flag possible rear-end collisions, then confirms with a YOLOv8 image detector across three consecutive frames and all cameras. If the annotations hold up, the dataset gives the community something synthetic accident sets cannot: real long-tail events at speed, with ground truth for detection, tracking, and prediction.

What carries the argument

The carrying mechanism is the Accid3nD pipeline: a trajectory-threshold rule module that scans vehicle tracks for rear-end collision signatures, a YOLOv8 image classifier that confirms the crash, a score threshold of 0.8, a three-frame consistency check, and a camera fusion step that merges detections from all roadside views. The dataset itself, with per-frame 2D/3D boxes and track IDs in OpenLABEL format, is the substrate that makes training and evaluation possible.

What would settle it

Take 1,000 randomly chosen frames from the dataset, have independent annotators re-label them with the same 3D tool, and measure overlap (IoU) between original and new boxes; if the average overlap falls below standard 3D detection thresholds, the ground-truth claim fails. Separately, run Accid3nD over long continuous roadside recordings with known crash times; if it misses every non-rear-end crash type, its claimed robustness is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that real-world highway accidents, not simulated ones, can serve as the training and evaluation ground for accident detection. The TUM Traffic Accident dataset records ten crash sequences at 10 Hz from roadside cameras and LiDARs on a German autobahn test stretch, delivering 294,924 2D and 93,012 3D annotations with track IDs across 48,144 labeled frames and ten object classes. On top of this data the paper builds Accid3nD, a two-stage detector: a rule-based trajectory module flags vehicles as accident candidates using predefined thresholds, and a learning-based YOLOv8 model then confirms the accident at image level. Detections must pass a score of 0.8 and appear in thre

Load-bearing premise

The load-bearing premise is that the published boxes and track IDs are accurate enough to serve as ground truth; the paper does not report any annotator-agreement or quality-control statistics.

Editorial extensions

If this is right

  • Accident detection can be trained and evaluated on real highway crash data rather than synthetic scenes, narrowing the sim-to-real gap.
  • Dense 3D boxes and track IDs support downstream tasks such as multi-object tracking, trajectory prediction, sensor fusion, and digital twin construction.
  • The three-frame confirmation and multi-camera fusion offer a practical recipe for reducing false positive accident alerts in roadside monitoring.
  • The rule-based stage's explicit limitation to rear-end collisions defines a clear baseline for extending accident detection to other collision types.
  • A real-time runtime of roughly 95 FPS suggests the pipeline could be deployed in continuous highway monitoring systems.

Reading between the lines

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

  • Beyond the paper: the recorded counts of standing vehicles in driving lanes and shoulder lanes could be repurposed as a large-scale near-miss or risk indicator dataset, not just crash data.
  • Beyond the paper: because no annotator-agreement statistics are reported, a natural independent check is to re-label a sample of frames and measure agreement; if agreement is high, the dataset becomes a benchmark for roadside infrastructure perception.
  • Beyond the paper: the hybrid rule-plus-learning design could be transferred to other roadside deployments without annotated crash data, using trajectory rules alone to generate pseudo-labels for a learned verifier.
  • Beyond the paper: the paper processes 128 days of continuous recordings but reports only aggregate counts; per-day false-alarm rates would be a direct, testable extension of the claimed robustness.
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

4 major / 5 minor

Summary. The paper introduces the TUM Traffic Accident (TUMTraf-A) dataset, a collection of real-world highway accidents recorded from roadside cameras and LiDARs, containing 294,924 2D and 93,012 3D annotations with track IDs across 48,144 labeled frames. It also proposes Accid3nD, a two-stage accident detection framework combining a rule-based trajectory analyzer with a learning-based YOLOv8 image classifier. The abstract and contributions claim state-of-the-art results based on experiments and ablation studies; however, the Evaluation section (Section V) reports only runtime performance, with no detection accuracy metrics, no baselines, and no ablations.

Significance. If the dataset annotations are reliable, TUMTraf-A is a potentially valuable resource: naturally occurring highway accidents are rare, and a dataset with dense 2D/3D boxes, track IDs, and multi-view roadside sensors could support research on long-tail safety events, perception, tracking, and cooperative sensing. The authors also share dataset, model, and code, which is a strength. However, the scientific claims about the Accid3nD detector are currently unsupported: the evaluation does not quantify detection accuracy, and the annotation quality is not demonstrated. The dataset contribution is significant but the paper as a whole needs substantial additional evidence to justify the stated conclusions.

major comments (4)
  1. [Section V] The 'Evaluation' section reports only runtime: 10.41 ms/frame (95.05 FPS) and 234.25 s for a 15-minute rosbag. There are no precision/recall values, F1 scores, mAP, accuracy, confusion matrices, or comparisons to existing accident-detection methods. The abstract and Section I claim 'state-of-the-art results' based on 'experiments and ablation studies,' but no such experiments or ablations appear anywhere in the manuscript. This is load-bearing: without detection-quality metrics, the reader cannot verify that Accid3nD detects accidents at all, let alone at state-of-the-art level. The paper must provide these numbers or substantially revise its claims.
  2. [Section III] The dataset is described as ground truth for perception tasks, but annotation quality is never quantified. The text states that labeling was done with the 3D BAT toolbox but gives no inter-annotator agreement, quality-control statistics, manual verification procedures, or error rates. For a dataset that is meant to support safety-critical perception, annotation accuracy is a central load-bearing property. If boxes or track IDs contain significant errors, both dataset value and any downstream evaluation are compromised. Please include annotation-quality measures (e.g., IAA on a subset, re-labeling statistics, or a comparison with existing datasets' QC practices).
  3. [Section VI] The Conclusion admits that the rule-based approach 'can only detect rear-end collisions.' Yet the abstract and contributions claim the method detects 'accidents and near-miss events' and achieves state-of-the-art results on a dataset containing overturns, fires, and various collision types. The claimed generality is therefore unsupported. Provide a breakdown of accuracy per accident type or scope the claims to rear-end collisions explicitly; otherwise the evaluation must cover all featured accident categories.
  4. [Section IV] The detection thresholds—YOLOv8 confidence score 0.8 and a minimum of three consecutive frames—are introduced without any ablation or sensitivity analysis. Since no accuracy evaluation is given, these hand-set parameters appear arbitrary, and the robustness claim ('the robustness of our proposed method') is ungrounded. The authors should either provide an ablation over thresholds or acknowledge that these values are not validated.
minor comments (5)
  1. [Affiliations] Typo: 'Laboratory for Intelligent and Safe Autombiles' should be 'Automobiles.'
  2. [Section II] The claim that 'Existing accident detection methods have never been tested on real traffic data of a test stretch' is a strong universal statement without supporting citation. Please soften or provide evidence.
  3. [Section IV] Training details for YOLOv8 (train/validation split, number of epochs, input resolution, data augmentation) are missing. Without these, the runtime result alone is not reproducible.
  4. [Section III/IV] The relationship between the 48,144 labeled frames and the 12,290 15-minute videos processed in Section IV is unclear. How many actual accident events are in the dataset, and how many of those are in the 48,144 frames? Please clarify.
  5. [Abstract] The abstract states 'Experiments and ablation studies show...' but no ablation study appears in the manuscript. Either add the ablation or remove the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper presents a dataset and a detector, and the cited self-citations concern tooling/lineage; the missing accuracy evaluation is an evidence gap, not a circular reduction.

full rationale

The paper contains no derivation chain that reduces a prediction to its inputs. Accid3nD combines a rule-based trajectory check with a YOLOv8 image classifier trained on the presented dataset, which is standard practice. The score threshold (0.8) and the three-frame requirement are hand-set, not fitted to an evaluation metric, so there is no fitted-input-called-prediction pattern. The main claim of state-of-the-art results is unsupported because Section V, titled 'Evaluation', reports only runtime performance (10.41 ms/frame, 95.05 FPS, 234.25 s per 15-minute rosbag) and gives no detection accuracy, mAP, precision/recall, or ablation table. That is a correctness/verifiability deficit, not circularity: the missing numbers would not be circular even if supplied, and their absence does not make any stated result equivalent to an input by construction. Self-citations to the 3D BAT annotation toolbox and to prior TUM Traffic datasets concern annotation tooling and dataset lineage; they are not load-bearing in any argument that derives a result from itself. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The main free parameters are hand-set detection thresholds. The method relies on the unstated availability of accurate vehicle trajectories and assumes that accidents of interest are visible from roadside sensors. No new entities are proposed.

free parameters (3)
  • rule-based detection thresholds
    Predefined thresholds for trajectory-based accident detection, values not reported in the paper, Section IV.
  • YOLOv8 confidence threshold = 0.8
    Score filter for accident detections, chosen by hand, Section IV.
  • consecutive frames minimum = 3
    An accident must be detected in at least three consecutive frames to reduce false positives, Section IV.
assumptions (3)
  • domain assumption Roadside cameras and LiDAR provide sufficient coverage to capture accidents
    The entire method and dataset depend on the sensors being positioned and functioning such that accidents are visible; no sensor coverage analysis is provided.
  • domain assumption Vehicle trajectories are available and accurate
    The rule-based detector operates on vehicle trajectories; the paper does not state how these trajectories are produced or their error characteristics, Section IV.
  • domain assumption All relevant accidents are of the rear-end type
    The conclusion states the rule-based approach can only detect rear-end collisions, which limits the dataset labeling and detection scope.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safety-Critical Learning for Long-Tail Events: The TUM Traffic Accident Dataset." pith.science (2026). https://pith.science/paper/YV2VGPHN

@misc{pith2026250814567,
  author       = {Pith},
  title        = {Pith review of: Safety-Critical Learning for Long-Tail Events: The TUM Traffic Accident Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YV2VGPHN}},
  note         = {Machine review of arXiv:2508.14567}
}
read the original abstract

Even though a significant amount of work has been done to increase the safety of transportation networks, accidents still occur regularly. They must be understood as an unavoidable and sporadic outcome of traffic networks. We present the TUM Traffic Accident (TUMTraf-A) dataset, a collection of real-world highway accidents. It contains ten sequences of vehicle crashes at high-speed driving with 294,924 labeled 2D and 93,012 labeled 3D boxes and track IDs within 48,144 labeled frames recorded from four roadside cameras and LiDARs at 10 Hz. The dataset contains ten object classes and is provided in the OpenLABEL format. We propose Accid3nD, an accident detection model that combines a rule-based approach with a learning-based one. Experiments and ablation studies on our dataset show the robustness of our proposed method. The dataset, model, and code are available on our project website: https://tum-traffic-dataset.github.io/tumtraf-a.

Figures

Figures reproduced from arXiv: 2508.14567 by the authors.

Figure 1
Figure 1. Visualization of the TUM Traffic Accident dataset with 3D box annotations, track IDs and trajectories. Accidents are recorded from four roadside cameras on the A9 Test Bed for Autonomous Driving in Munich, Germany. Left: a vehicle is in the process of overturning following collision. Right: a vehicle has pulled over after catching fire following a collision. Abstract—Even though a significant amount of work has been… view at source ↗
Figure 2
Figure 2. Qualitative visualization results of our accident detection framework on the TUM Traffic Accident test set. Left: The rule-based approach detected a rear-end collision. Right: The learning-based approach detected a car crash. II. RELATED WORK Existing accident detection methods have never been tested on real traffic data of a test stretch. Real accident datasets are rare and do not contain enough data to train deep … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages

  1. [1]

    Planning with occluded traffic agents using bi-level variational occlu- sion models,

    F. Christianos, P. Karkus, B. Ivanovic, S. V . Albrecht, and M. Pavone, “Planning with occluded traffic agents using bi-level variational occlu- sion models,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 5558–5565, IEEE, 2023

    work page 2023

  2. [2]

    Activeanno3d-an active learning framework for multi-modal 3d object detection,

    A. Ghita, B. Antoniussen, W. Zimmer, R. Greer, C. Creß, A. Møgelmose, M. Trivedi, and A. C. Knoll, “Activeanno3d-an active learning framework for multi-modal 3d object detection,” in 35th IEEE Intelligent Vehicles Symposium (IV) 2024 , 2024

    work page 2024

  3. [3]

    Create a large-scale video driving dataset with detailed attributes using amazon sagemaker ground truth,

    N. Kulkarni, A. Rangesh, J. Buck, J. Feltracco, M. Trivedi, N. Deo, R. Greer, S. Sarraf, and S. Sathyanarayana, “Create a large-scale video driving dataset with detailed attributes using amazon sagemaker ground truth,” 2021

    work page 2021

  4. [4]

    Fingscheidt, H

    T. Fingscheidt, H. Gottschalk, and S. Houben, Deep neural networks and data for automated driving: Robustness, uncertainty quantification, and insights towards safety . Springer Nature, 2022

    work page 2022

  5. [5]

    Drive video analysis for the detection of traffic near-miss incidents,

    H. Kataoka, T. Suzuki, S. Oikawa, Y . Matsui, and Y . Satoh, “Drive video analysis for the detection of traffic near-miss incidents,” in 2018 IEEE International Conference on robotics and automation (ICRA) , pp. 3421–3428, IEEE, 2018

    work page 2018

  6. [6]

    Ips300+: a challenging multi-modal data sets for intersection per- ception system,

    H. Wang, X. Zhang, Z. Li, J. Li, K. Wang, Z. Lei, and R. Haibing, “Ips300+: a challenging multi-modal data sets for intersection per- ception system,” in 2022 International Conference on Robotics and Automation (ICRA), pp. 2539–2545, IEEE, 2022

    work page 2022

  7. [7]

    The Why, When, and How to Use Active Learning in Large-Data-Driven 3D Object Detection for Safe Autonomous Driving: An Empirical Exploration

    R. Greer, B. Antoniussen, M. V . Andersen, A. Møgelmose, and M. M. Trivedi, “The why, when, and how to use active learning in large-data- driven 3d object detection for safe autonomous driving: An empirical exploration,” arXiv preprint arXiv:2401.16634 , 2024

    work page Pith review arXiv 2024

  8. [8]

    GraphRelate3D: Context-Dependent 3D Object Detection with Inter-Object Relationship Graphs

    M. Liu, E. Yurtsever, M. Brede, J. Meng, W. Zimmer, X. Zhou, B. L. Zagar, Y . Cui, and A. Knoll, “Graphrelate3d: Context-dependent 3d object detection with inter-object relationship graphs,” arXiv preprint arXiv:2405.06782, 2024

    work page Pith review arXiv 2024

Show all 29 references
  1. [9]

    Roadsense3d: A framework for roadside monocular 3d object detection,

    S. Carta, M. Castrill ´on-Santana, M. Marras, S. Mohamed, A. S. Podda, R. Saia, M. Sau, and W. Zimmer, “Roadsense3d: A framework for roadside monocular 3d object detection,” in Adjunct Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Personalization, pp....

    2024

  2. [10]

    Infradet3d: Multi-modal 3d object de- tection based on roadside infrastructure camera and lidar sensors,

    W. Zimmer, J. Birkner, M. Brucker, H. T. Nguyen, S. Petrovski, B. Wang, and A. C. Knoll, “Infradet3d: Multi-modal 3d object de- tection based on roadside infrastructure camera and lidar sensors,” in 2023 IEEE Intelligent Vehicles Symposium (IV) , IEEE, 2023

    2023

  3. [11]

    Real-time and robust 3d object detection with roadside lidars,

    W. Zimmer, J. Wu, X. Zhou, and A. C. Knoll, “Real-time and robust 3d object detection with roadside lidars,” in Proc. of Int. Scientific Conf. on Mobility and Transport: Mobility Innovations for Growing Megacities, pp. 199–219, Springer, 2023

    2023

  4. [12]

    A survey of robust 3d object detection methods in point clouds,

    W. Zimmer, E. Ercelik, X. Zhou, X. J. D. Ortiz, and A. Knoll, “A survey of robust 3d object detection methods in point clouds,” arXiv preprint arXiv:2204.00106, 2022

    2022 arXiv

  5. [13]

    Real-time and robust 3d object detection within road-side lidars using domain adaptation,

    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

  6. [14]

    Traffic light detection: A learning algorithm and evaluations on challenging dataset,

    M. P. Philipsen, M. B. Jensen, A. Møgelmose, T. B. Moeslund, and M. M. Trivedi, “Traffic light detection: A learning algorithm and evaluations on challenging dataset,” in 2015 IEEE 18th International Conference on Intelligent Transportation Systems , pp. 2341–2345, IEEE, 2015

    2015

  7. [15]

    Laneaf: Robust multi-lane detection with affinity fields,

    H. Abualsaud, S. Liu, D. B. Lu, K. Situ, A. Rangesh, and M. M. Trivedi, “Laneaf: Robust multi-lane detection with affinity fields,” IEEE Robotics and Automation Letters , vol. 6, no. 4, pp. 7477–7484, 2021

    2021

  8. [16]

    Patterns of vehicle lights: Addressing complexities of camera-based vehicle light datasets and metrics,

    R. Greer, A. Gopalkrishnan, M. Keskar, and M. M. Trivedi, “Patterns of vehicle lights: Addressing complexities of camera-based vehicle light datasets and metrics,” Pattern Recognition Letters, vol. 178, pp. 209– 215, 2024

    2024

  9. [17]

    A digital twin for teleoper- ation of vehicles in urban environments,

    P. Kremer, N. Nourani-Vatani, and S. Park, “A digital twin for teleoper- ation of vehicles in urban environments,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) , pp. 12521–12527, IEEE, 2023

    2023

  10. [18]

    Safe control transitions: Machine vision based observable readiness index and data-driven takeover time prediction,

    R. Greer, N. Deo, A. Rangesh, M. Trivedi, and P. Gunaratne, “Safe control transitions: Machine vision based observable readiness index and data-driven takeover time prediction,” in 27th International Tech- nical Conference on the Enhanced Safety of Vehicles (ESV) National High...

    2023

  11. [19]

    A survey on autonomous driving datasets: Statistics, annotation quality, and a future outlook,

    M. Liu, E. Yurtsever, J. Fossaert, X. Zhou, W. Zimmer, Y . Cui, B. L. Zagar, and A. C. Knoll, “A survey on autonomous driving datasets: Statistics, annotation quality, and a future outlook,” IEEE Transactions on Intelligent Vehicles, 2024

    2024

  12. [20]

    Deepaccident: A motion and accident prediction benchmark for v2x autonomous driving,

    T. Wang, S. Kim, J. Wenxuan, E. Xie, C. Ge, J. Chen, Z. Li, and P. Luo, “Deepaccident: A motion and accident prediction benchmark for v2x autonomous driving,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, pp. 5599–5606, 2024

    2024

  13. [21]

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

    C. Creß, W. Zimmer, L. Strand, M. Fortkord, S. Dai, V . Lakshmi- narasimhan, and A. Knoll, “A9-dataset: Multi-sensor infrastructure- based dataset for mobility research,” in 2022 IEEE Intelligent Vehicles Symposium (IV), pp. 965–970, IEEE, 2022

    2022

  14. [22]

    Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception,

    W. Zimmer, C. Creß, H. T. Nguyen, and A. C. Knoll, “Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception,” in 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) , pp. 1030–1037, IEEE, 2023

    2023

  15. [23]

    Tumtraf event: Calibration and fusion resulting in a dataset for roadside event-based and rgb cameras,

    C. Creß, W. Zimmer, N. Purschke, B. N. Doan, S. Kirchner, V . Laksh- minarasimhan, L. Strand, and A. C. Knoll, “Tumtraf event: Calibration and fusion resulting in a dataset for roadside event-based and rgb cameras,” IEEE Transactions on Intelligent Vehicles , 2024

    2024

  16. [24]

    Tumtraf v2x cooperative perception dataset,

    W. Zimmer, G. A. Wardana, S. Sritharan, X. Zhou, R. Song, and A. C. Knoll, “Tumtraf v2x cooperative perception dataset,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pp. 22668–22677, 2024

    2024

  17. [25]

    3d bat: A semi-automatic, web-based 3d annotation toolbox for full-surround, multi-modal data streams,

    W. Zimmer, A. Rangesh, and M. Trivedi, “3d bat: A semi-automatic, web-based 3d annotation toolbox for full-surround, multi-modal data streams,” in 2019 IEEE Intelligent Vehicles Symposium (IV), pp. 1816– 1821, IEEE, 2019

    2019

  18. [26]

    Tum traffic datasets

    W. Zimmer, C. Creß, X. Zhou, L. Strand, V . Lakshminarasimhan, and A. Knoll, “Tum traffic datasets.” https://innovation-mobility.com/ tumtraf-dataset

  19. [27]

    TUM traf- fic dataset development kit

    W. Zimmer, C. Creß, X. Zhou, and A. Knoll, “TUM traf- fic dataset development kit.” https://github.com/tum-traffic-dataset/ tum-traffic-dataset-dev-kit

  20. [28]

    Towards explainable, safe autonomous driving with language embeddings for novelty identification and active learning: Framework and experimental analysis with real-world data sets,

    R. Greer and M. Trivedi, “Towards explainable, safe autonomous driving with language embeddings for novelty identification and active learning: Framework and experimental analysis with real-world data sets,” arXiv preprint arXiv:2402.07320 , 2024

    2024 arXiv

  21. [29]

    Pedestrian behavior maps for safety advisories: Champ framework and real-world data analysis,

    R. Greer, S. Desai, L. Rakla, A. Gopalkrishnan, A. Alofi, and M. Trivedi, “Pedestrian behavior maps for safety advisories: Champ framework and real-world data analysis,” in 2023 IEEE Intelligent Vehicles Symposium (IV), pp. 1–8, IEEE, 2023

    2023

Pith tools

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