Pith. sign in

REVIEW 3 major objections 4 minor 32 references

A Point Cloud Completion Approach for the Grasping of Partially Occluded Objects and Its Applications in Robotic Strawberry Harvesting

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read This paper claims that completing partially occluded strawberry point clouds before grasp planning raises whole-fruit harvesting success rates and sharply reduces collisions with neighboring fruit.

desk verdict A competent integrated system for strawberry harvesting with real-robot numbers, but the evaluation doesn't prove the completion module is the cause of the improvement. read the letter →

arxiv 2506.14066 v1 pith:36OH5T3K submitted 2025-06-16 cs.RO

classification cs.RO
keywords pointcloudcompletionroboticstrawberryharvestingocclusionhandlinggraspplanningobstacleavoidanceoccupancymapChamferDistanceRGB-Dperception
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 argues that the missing piece in robotic strawberry harvesting is not detection alone but geometric reconstruction under occlusion. Its pipeline takes a single RGB-D view of a plant, isolates each strawberry, completes the partial point cloud of every fruit, and then plans a grasp toward the nearest ripe strawberry while treating all other strawberries as obstacles in an occupancy map. The claim, supported by 80 real-world trials, is that this combination yields a grasp success rate of 79.17%, a success-to-attempt ratio of 89.58%, a reconstruction error of 1.10 mm Chamfer Distance, and a drop in obstacle hit rate from 43.33% to 13.95%. A careful reader would care because it offers a concrete route to damage-free automated picking of a fragile crop in cluttered scenes.

What carries the argument

The load-bearing object is the point cloud completion network, an attention-based model following PointAttn, which maps a denoised partial point cloud of a strawberry to progressively denser completed point clouds under a hierarchical Chamfer Distance loss. It is paired with an occupancy-map planner: the completed point clouds of non-target strawberries are unioned and discretized into a 3D occupancy grid, so inverse kinematics and trajectory planning treat neighboring fruit as obstacles while approaching the selected ripe target.

What would settle it

Run the same pipeline on strawberries whose sizes span a realistic range (for example, plus or minus 30% of the fixed CAD dimensions) in the same lab setup and measure the grasp success rate and Chamfer Distance. If the success-to-attempt ratio falls back toward the 74.07% baseline, or the completed point clouds deviate by more than about 2 mm once size deviates, the fixed-shape assumption is the weak link and the central claim loses its generality.

Watch

Extended reading notes

Core claim

The central claim is that a complete 3D reconstruction of every strawberry in a scene, not just the target, is what enables safe whole-fruit grasping under occlusion. For each detected strawberry the pipeline produces a dense completed point cloud from a noisy partial scan, selects the ripe strawberry closest to the end-effector, and builds an occupancy map from the completed point clouds of all other strawberries so the planned trajectory avoids them. In real-world trials this pipeline achieves a grasping success rate of 79.17% with a success-to-attempt ratio of 89.58%, a Chamfer Distance of 1.10 mm that is lower than the compared completion baselines, and a reduction in obstacle hit rate from 43.33% to 13.95% when the obstacle-aware planning is enabled.

Load-bearing premise

The pipeline assumes every strawberry has one fixed known shape and size from a CAD model, so if real fruit vary in size, shape, or ripeness deformation, the completed point clouds and the grasps built from them become inaccurate.

Editorial extensions

If this is right

  • In cluttered scenes, completing the geometry of both target and non-target fruit reduces collisions; the reported obstacle hit rate drops by about 68% when the full pipeline is used.
  • The lower Chamfer Distance (1.10 mm vs 2.42 mm for the best baseline) translates into more reliable grasp localization, which is the stated reason for the higher grasping success rate.
  • Treating unripe and occluded strawberries as obstacles rather than discarding them allows the robot to harvest without damaging the surrounding crop.
  • Because the pipeline works from a single RGB-D view on an embedded GPU, it is compatible with field-style deployment conditions that allow only one sensor pass.
  • The success-to-attempt ratio of 89.58% compared to 74.07% for the shape-completion baseline indicates fewer failed attempts per successful harvest, which matters for cycle time in a real picking operation.

Reading between the lines

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

  • If the central claim holds, then the same pattern of completing all objects in a scene before grasp planning should transfer to other roughly rigid fruits such as apples or tomatoes, provided the completion model is retrained on their shapes.
  • A natural ablation test would separate the two contributions: planning on partial point clouds with obstacle avoidance would isolate how much of the gain comes from completion itself, while planning on completed clouds without the occupancy map would isolate the obstacle-avoidance contribution.
  • The reported drop in obstacle hit rate suggests that completion improves obstacle localization as well as target reconstruction; quantifying those two effects separately would sharpen the claim.
  • Field deployment is the real test: the fixed-shape CAD assumption and lab lighting mean the reported numbers may not represent natural variability in berry size, ripeness deformation, and leaf clutter.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes an end-to-end robotic strawberry harvesting pipeline that combines point cloud denoising/segmentation, PointAttn-based point cloud completion, target selection based on ripeness and distance, and occupancy-map-aware motion planning that treats non-target strawberries as obstacles. The authors report a Chamfer Distance of 1.10 mm for shape completion, a grasp success rate of 79.17% with a success-to-attempt ratio of 89.58%, and a reduction in obstacle hit rate from 43.33% to 13.95% over a baseline. The evaluation is based on 80 real-world lab trials with five strawberries and a Unitree Z1 arm, with code publicly available.

Significance. If the claims hold, the paper makes a useful contribution to agricultural robotics by demonstrating that combining point cloud completion with obstacle-aware planning can improve grasping of partially occluded fruit. The strengths are the public code release, a real-robot evaluation, and a clean pipeline description that others can reproduce. However, the significance is currently conditional: the grasping comparison uses a different robotic platform and environment, the ablation in Table III does not isolate the completion module, the Chamfer Distance evaluation lacks a clear common protocol, and no statistical backing is provided for the claimed improvements. The stated assumption of a fixed strawberry shape and size further limits the generality of the results, although the authors acknowledge this in the future-work section.

major comments (3)
  1. [§V.E, Table I] The only grasping baseline, Shape Completion [9], was evaluated on a different robot, gripper, camera, and vertical-farm environment, which the authors acknowledge. Consequently, the reported 26.67% relative improvement in ρs is not attributable to the proposed method alone. A same-platform baseline or an ablation of the proposed planning/completion components on the authors' own robot is needed to support the central claim that the pipeline improves grasping.
  2. [§VI, Table III] The 'Ours Used ×' row does not specify which components are disabled. If the × condition removes only the obstacle-aware planner while point cloud completion remains active, then the reduction in obstacle hit rate from 43.33% to 13.95% provides no evidence about the contribution of completion to grasping, and completion is never ablated in the grasping evaluation. The paper must define the ablated configuration precisely and ideally include separate ablations for completion and obstacle-aware planning. In addition, no raw trial counts, confidence intervals, or statistical tests are reported for the 80 trials, so the word 'significantly' in the abstract and Section VI is unsupported.
  3. [§V.A, §VI, Table II] The Chamfer Distance comparison lacks a common evaluation protocol. It is not stated whether the baselines [24] and [32] were run on the same partial inputs, the same ground truth, and the same evaluation split as the proposed method. Since the lab-collected real data were manually annotated, combined with simulation data for training, and also used for the reported CD evaluation, the paper must disclose whether the strawberries used for the CD test overlap with the training set; otherwise the 1.10 mm result may reflect memorized shapes rather than genuine completion capability.
minor comments (4)
  1. [§IV.A, Eqs. (4)-(5)] The voxel size vs and the minimum-points-per-voxel threshold of 30 are given without units and without justification. Also, the set-builder notation 'pj∈ p| floor(pj/vs)=i' is malformed; please write the condition over voxel indices explicitly.
  2. [§V.C] The detection and segmentation modules YOLOv8 and SAM2 are described only by name; no detection or segmentation accuracy metrics are reported, and the paper does not explain how mask-point association is performed when multiple masks overlap or when detection misses a strawberry. This makes it hard to assess how perception errors propagate to grasp success.
  3. [Algorithm 1, lines 14-17] The occupancy map Mocc is generated only from obstacle point clouds, while the target's completed point cloud is used for grasp estimation. Please clarify how the target geometry is kept out of the collision map and how the completed target shape influences the planned trajectory.
  4. [Tables I and III] The denominators for ρa, ρs, and ρh are not explicitly stated. Reporting the raw counts (number of detections, attempts, successes, and obstacle hits) would allow the rates to be checked and would partially mitigate the absence of confidence intervals.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the pipeline is an empirical system evaluation with no derivation that reduces to its inputs.

full rationale

This paper does not contain a derivation chain in which an output is equivalent to its inputs by construction. The central claims are experimental measurements: grasp success rates, Chamfer Distance, and obstacle hit rates obtained from a robotic pipeline integrating detection, segmentation, point cloud completion, and collision-aware planning. Point cloud completion is implemented via a learned network based on an external architecture (PointAttn), with pre-trained weights trained on simulated and real data; this is an empirical fitting procedure, not a first-principles prediction, and the paper does not rename a fitted parameter as a predicted outcome. The comparisons against Shape Completion [9], CoRe [24], and DeepSDF-based [32] are external benchmarks with no author overlap, so no load-bearing self-citation chain is present. The stated fixed-CAD-shape assumption is explicit and is a modeling limitation, not a circular step. The only potential validity concern is that the manuscript does not clearly describe a train/test split between the real-world data used for training the completion model and the 80 real-world evaluation trials; if the evaluation scenes overlapped the training data, that would be an experimental leakage issue, but the paper provides no quoted reduction showing that the reported predictions are forced by construction from the training data. Under the specified circularity criteria, no circular step can be exhibited, and the appropriate finding is no significant circularity.

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

The pipeline introduces no new physical entities and no new mathematical derivation. Its central assumptions are the fixed-shape strawberry model and the fidelity of the pretrained completion network.

free parameters (4)
  • Voxel size vs = 0.05
    Used in Eq. 4 for downsampling; chosen by hand.
  • Minimum points per voxel threshold = 30
    Used in Eq. 5 to filter noise; chosen by hand.
  • Median blur window size = 5x5
    Used in Eq. 1 for depth filtering; chosen by hand.
  • Completion loss weights lambda_i = not reported
    Weights in Eq. 11 for multi-resolution CD loss; values are not disclosed.
assumptions (4)
  • domain assumption Strawberry shape and size are represented by a fixed CAD model (Section III).
    The completion model is trained and evaluated under the assumption that strawberries have a fixed shape, which limits generalization to variable fruit.
  • domain assumption The pre-trained PointAttn model generalizes to strawberry point clouds after fine-tuning (Section IV-B).
    The completion network is taken from [7] and assumed to work for this domain.
  • domain assumption The occupancy map built from completed point clouds is accurate enough for collision-free planning (Section IV-C).
    Planning relies on the completed obstacle point clouds being geometrically faithful.
  • standard math Camera intrinsics K are known and depth projection is error-free (Section IV-A).
    The point cloud generation uses the standard pinhole projection, which is standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Point Cloud Completion Approach for the Grasping of Partially Occluded Objects and Its Applications in Robotic Strawberry Harvesting." pith.science (2026). https://pith.science/paper/36OH5T3K

@misc{pith2026250614066,
  author       = {Pith},
  title        = {Pith review of: A Point Cloud Completion Approach for the Grasping of Partially Occluded Objects and Its Applications in Robotic Strawberry Harvesting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36OH5T3K}},
  note         = {Machine review of arXiv:2506.14066}
}
read the original abstract

In robotic fruit picking applications, managing object occlusion in unstructured settings poses a substantial challenge for designing grasping algorithms. Using strawberry harvesting as a case study, we present an end-to-end framework for effective object detection, segmentation, and grasp planning to tackle this issue caused by partially occluded objects. Our strategy begins with point cloud denoising and segmentation to accurately locate fruits. To compensate for incomplete scans due to occlusion, we apply a point cloud completion model to create a dense 3D reconstruction of the strawberries. The target selection focuses on ripe strawberries while categorizing others as obstacles, followed by converting the refined point cloud into an occupancy map for collision-aware motion planning. Our experimental results demonstrate high shape reconstruction accuracy, with the lowest Chamfer Distance compared to state-of-the-art methods with 1.10 mm, and significantly improved grasp success rates of 79.17%, yielding an overall success-to-attempt ratio of 89.58\% in real-world strawberry harvesting. Additionally, our method reduces the obstacle hit rate from 43.33% to 13.95%, highlighting its effectiveness in improving both grasp quality and safety compared to prior approaches. This pipeline substantially improves autonomous strawberry harvesting, advancing more efficient and reliable robotic fruit picking systems.

Figures

Figures reproduced from arXiv: 2506.14066 by the authors.

Figure 1
Figure 1. Top: Input RGB-D image pair showing a multiple strawberry [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our robotic strawberry harvesting pipeline consists of: Point Cloud Denoising and Segmentation, Point Cloud Completion [7], and Target Selection [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Simulation and real-world environments for data collection. (Left) Real-world lab setup replicating an indoor greenhouse strawberry plantation. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Robotic system for strawberry harvesting. The image shows the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [9]

    Improving robotic fruit harvesting within cluttered environments through 3d shape completion,

    F. Magistri, Y . Pan, J. Bartels, J. Behley, C. Stachniss, and C. Lehnert, “Improving robotic fruit harvesting within cluttered environments through 3d shape completion,” IEEE Robotics and Automation Letters , 2024

  2. [24]

    Contrastive 3d shape completion and reconstruction for agricultural robots using rgb-d frames,

    F. Magistri, E. Marks, S. Nagulavancha, I. Vizzo, T. L ¨aebe, J. Behley, M. Halstead, C. McCool, and C. Stachniss, “Contrastive 3d shape completion and reconstruction for agricultural robots using rgb-d frames,” IEEE Robotics and Automation Letters , vol. 7, no. 4, pp. 10 120–10 127, 2022

  3. [32]

    Panoptic mapping with fruit completion and pose estimation for horticultural robots,

    Y . Pan, F. Magistri, T. L¨abe, E. Marks, C. Smitt, C. McCool, J. Behley, and C. Stachniss, “Panoptic mapping with fruit completion and pose estimation for horticultural robots,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 4226–4233

  4. [1]

    Tech- nology progress in mechanical harvest of fresh market strawberries,

    J. Guo, Z. Yang, M. Karkee, Q. Jiang, X. Feng, and Y . He, “Tech- nology progress in mechanical harvest of fresh market strawberries,” Computers and Electronics in Agriculture , vol. 226, p. 109468, 2024

  5. [2]

    Labour and skills shortages in the agro-food sector,

    M. Ryan, “Labour and skills shortages in the agro-food sector,” OECD F ood, Agriculture and Fisheries Papers, no. 189, 2023

  6. [3]

    A review of perception technologies for berry fruit-picking robots: Advantages, disadvantages, challenges, and prospects,

    C. Wang, W. Pan, T. Zou, C. Li, Q. Han, H. Wang, J. Yang, and X. Zou, “A review of perception technologies for berry fruit-picking robots: Advantages, disadvantages, challenges, and prospects,” Agriculture, vol. 14, no. 8, p. 1346, 2024

  7. [4]

    Intelligent robots for fruit harvesting: Recent developments and future challenges,

    H. Zhou, X. Wang, W. Au, H. Kang, and C. Chen, “Intelligent robots for fruit harvesting: Recent developments and future challenges,” Precision Agriculture, vol. 23, no. 5, pp. 1856–1907, 2022

  8. [5]

    Sim-to-real grasp detection with global-to-local rgb-d adaptation,

    H. Ma, R. Qin, M. Shi, B. Gao, and D. Huang, “Sim-to-real grasp detection with global-to-local rgb-d adaptation,” in 2024 IEEE Inter- national Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 13 910–13 917

Show all 32 references
  1. [6]

    Addressing the sim2real gap in robotic 3-d object classification,

    J.-B. Weibel, T. Patten, and M. Vincze, “Addressing the sim2real gap in robotic 3-d object classification,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 407–413, 2019

  2. [7]

    Pointattn: You only need attention for point cloud completion,

    J. Wang, Y . Cui, D. Guo, J. Li, Q. Liu, and C. Shen, “Pointattn: You only need attention for point cloud completion,” 2022. [Online]. Available: https://arxiv.org/abs/2203.08485

  3. [8]

    Peduncle gripping and cutting force for strawberry harvesting robotic end-effector design,

    S. Parsa, S. Parsons, A. Ghalamzan et al. , “Peduncle gripping and cutting force for strawberry harvesting robotic end-effector design,” in 2022 4th International Conference on Control and Robotics (ICCR) . IEEE, 2022, pp. 59–64

  4. [10]

    Modular autonomous strawberry picking robotic system,

    S. Parsa, B. Debnath, M. A. Khan, and A. G. E, “Modular autonomous strawberry picking robotic system,” Journal of Field Robotics , vol. 41, no. 7, pp. 2226–2246, 2024

  5. [11]

    A berry picking robot with a hybrid soft-rigid arm: Design and task space control

    N. K. Uppalapati, B. Walt, A. J. Havens, A. Mahdian, G. Chowdhary, and G. Krishnan, “A berry picking robot with a hybrid soft-rigid arm: Design and task space control.” in Robotics: Science and Systems , 2020, p. 95

  6. [12]

    State-of- the-art robotic grippers, grasping and control strategies, as well as their applications in agricultural robots: A review,

    B. Zhang, Y . Xie, J. Zhou, K. Wang, and Z. Zhang, “State-of- the-art robotic grippers, grasping and control strategies, as well as their applications in agricultural robots: A review,” Computers and Electronics in Agriculture , vol. 177, p. 105694, 2020

  7. [13]

    Safe leaf manipulation for accurate shape and pose estimation of occluded fruits,

    S. Yao, S. Pan, M. Bennewitz, and K. Hauser, “Safe leaf manipulation for accurate shape and pose estimation of occluded fruits,” arXiv preprint arXiv:2409.17389, 2024

  8. [14]

    Occlusion handling by pushing for enhanced fruit detection,

    E. Gursoy, D. Kuli ´c, and A. Cherubini, “Occlusion handling by pushing for enhanced fruit detection,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 76–81

  9. [15]

    Peduncle detection of sweet pepper for autonomous crop harvesting—combined color and 3-d information,

    I. Sa, C. Lehnert, A. English, C. McCool, F. Dayoub, B. Upcroft, and T. Perez, “Peduncle detection of sweet pepper for autonomous crop harvesting—combined color and 3-d information,” IEEE Robotics and Automation Letters, vol. 2, no. 2, pp. 765–772, 2017

  10. [16]

    Efficient and robust orientation estimation of strawberries for fruit picking appli- cations,

    N. Wagner, R. Kirk, M. Hanheide, and G. Cielniak, “Efficient and robust orientation estimation of strawberries for fruit picking appli- cations,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 13 857–13 863

  11. [17]

    Single-shot 6dof pose and 3d size estimation for robotic strawberry harvesting,

    L. Li and H. Kasaei, “Single-shot 6dof pose and 3d size estimation for robotic strawberry harvesting,” 2024. [Online]. Available: https://arxiv.org/abs/2410.03031

  12. [18]

    Key point-based orientation es- timation of strawberries for robotic fruit picking,

    J. Le Lou ¨edec and G. Cielniak, “Key point-based orientation es- timation of strawberries for robotic fruit picking,” in International Conference on Computer Vision Systems . Springer, 2023, pp. 148– 158

  13. [19]

    A mango picking vision algorithm on instance segmentation and key point detection from rgb images in an open orchard,

    C. Zheng, P. Chen, J. Pang, X. Yang, C. Chen, S. Tu, and Y . Xue, “A mango picking vision algorithm on instance segmentation and key point detection from rgb images in an open orchard,” Biosystems engineering, vol. 206, pp. 32–54, 2021

  14. [20]

    Straw- berry picking point localization ripeness and weight estimation,

    A. Tafuro, A. Adewumi, S. Parsa, G. E. Amir, and B. Debnath, “Straw- berry picking point localization ripeness and weight estimation,” in 2022 International conference on robotics and automation (ICRA) . Ieee, 2022, pp. 2295–2302

  15. [21]

    Deepsdf: Learning continuous signed distance functions for shape representation,

    J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove, “Deepsdf: Learning continuous signed distance functions for shape representation,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2019, pp. 165–174

  16. [22]

    Learning a structured latent space for unsupervised point cloud completion,

    Y . Cai, K.-Y . Lin, C. Zhang, Q. Wang, X. Wang, and H. Li, “Learning a structured latent space for unsupervised point cloud completion,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5543–5553

  17. [23]

    Mv- deepsdf: Implicit modeling with multi-sweep point clouds for 3d vehicle reconstruction in autonomous driving,

    Y . Liu, K. Zhu, G. Wu, Y . Ren, B. Liu, Y . Liu, and J. Shan, “Mv- deepsdf: Implicit modeling with multi-sweep point clouds for 3d vehicle reconstruction in autonomous driving,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 8306–8316

  18. [25]

    Sweet pepper pose detection and grasping for automated crop harvesting,

    C. F. Lehnert, I. Sa, C. McCool, B. Upcroft, and T. Perez, “Sweet pepper pose detection and grasping for automated crop harvesting,” 2016 IEEE International Conference on Robotics and Automation (ICRA) , pp. 2428–2434, 2016. [Online]. Available: https://api.semanticscholar.org...

  19. [26]

    Mobile robotics platform for strawberry sensing and harvesting within precision indoor farming systems,

    G. Ren, T. Wu, T. Lin, L. Yang, G. Chowdhary, K. Ting, and Y . Ying, “Mobile robotics platform for strawberry sensing and harvesting within precision indoor farming systems,” J. Field Robotics, vol. 41, pp. 2047–2065, 2023. [Online]. Available: https://api.semanticscholar.org/...

  20. [27]

    Real-to-sim grasp: Rethinking the gap between simulation and real world in grasp detection,

    J.-F. Cai, Z. Chen, X.-M. Wu, J.-J. Jiang, Y .-L. Wei, and W.-S. Zheng, “Real-to-sim grasp: Rethinking the gap between simulation and real world in grasp detection,” arXiv preprint arXiv:2410.06521 , 2024

  21. [28]

    Bridging the simulation-to-real gap of depth images for deep reinforcement learning,

    Y . Jang, J. Baek, S. Jeon, and S. Han, “Bridging the simulation-to-real gap of depth images for deep reinforcement learning,” Expert Systems with Applications , vol. 253, p. 124310, 2024

  22. [29]

    Linear attention based deep nonlocal means filtering for multiplicative noise removal,

    X. Siyao, H. Libing, and Z. Shunsheng, “Linear attention based deep nonlocal means filtering for multiplicative noise removal,” 2024. [Online]. Available: https://arxiv.org/abs/2407.05087

  23. [30]

    Ultralytics yolov8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolov8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

  24. [31]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,” 2024. [Online]. Avail...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.