Pith. sign in

REVIEW 4 major objections 6 minor 47 references

PlantSAM: An Object Detection-Driven Segmentation Pipeline for Herbarium Specimens

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Object detection prompts SAM2 to isolate plants on herbarium scans.

desk verdict Useful engineering result, but the headline numbers mainly measure agreement with the authors' own prior pipeline, so the SOTA claim needs independent ground truth before it can be trusted. read the letter →

arxiv 2507.16506 v1 pith:P76D4AO6 submitted 2025-07-22 cs.CV

Youcef Sklab , Florian Castanet , Hanane Ariouat , Souhila Arib , Jean-Daniel Zucker , Eric Chenin , Edi Prifti This is my paper
classification cs.CV
keywords herbariumsegmentationSAM2YOLOv10promptablebackgroundremovalbotanicaltraitclassificationobjectdetectiondigitizedspecimens
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

PlantSAM is an automatic pipeline for stripping away the non-plant background of digitized herbarium sheets before any downstream analysis. The paper's thesis is that background clutter is a fixable bottleneck: if you detect where the plant is and feed that location to a fine-tuned Segment Anything Model, the resulting masks are accurate enough to improve morphological trait classification. On a 333-image test set the pipeline reports IoU 0.94 and Dice 0.97, beating a UNet baseline and a SAM1 variant, and it produces usable masks more often than UNet on colored backgrounds, thin stems, and pinned specimens. When the masks are used to crop images, ResNet101 classification accuracy rises by up to 4.36% and F1 by up to 4.15% across five botanical traits. If the numbers survive independent ground truth, this is a practical preprocessing tool for large herbarium collections.

What carries the argument

The load-bearing mechanism is detector-prompted segmentation. Large scans are split into patches; YOLOv10 proposes bounding boxes using a multi-region strategy that follows connected foreground clusters; those boxes are SAM2's only prompts; and the predicted patches are stitched back into a complete mask. The paper quantifies why the prompt matters: the multi-region strategy keeps 51.96% of the box area on plant material, versus 42.01% for a single enclosing box, so the detection step does much of the work of avoiding background noise.

What would settle it

Have independent annotators re-trace a random subset of the 333 test images by hand and compute PlantSAM2's IoU and Dice against those manual traces; if the scores fall materially below 0.94 and 0.97, the headline numbers partly reflect agreement with the semi-automatic reference masks. A second check is to rerun the trait-classification experiment using crops from manually drawn boxes rather than YOLO boxes and see whether the 4.36% gain persists.

Watch

Extended reading notes

Core claim

On its own terms, the paper shows that object detection can replace manual prompt design for promptable segmentation in a specialized domain. A fine-tuned YOLOv10 detector localizes plant material in image patches and emits bounding-box prompts; a fine-tuned SAM2 converts each box into a mask; and the masks are recombined into a full-resolution image mask. The authors claim this two-stage automatic prompting outperforms both a standalone SAM2 and a conventional UNet on herbarium images, with a multi-region prompting strategy that isolates connected components of the plant rather than one box per patch. They add a caveat that their own masks can exceed the annotation quality of the reference labels, so the IoU numbers may understate visual quality.

Load-bearing premise

The evaluation assumes that the masks from the earlier semi-automatic pipeline are accurate enough to be ground truth, because those masks generated YOLOv10's training boxes, the curated SAM2 fine-tuning set, and the 333-image test set; if those masks carry systematic errors, the IoU and Dice scores mostly show agreement with that pipeline, not true segmentation quality.

Editorial extensions

If this is right

  • Background removal before trait classification yields consistent gains across all five tested traits, with the largest observed on armatures: +4.36% accuracy.
  • Cropping to segmented plant regions preserves resolution, so fine morphological details survive the fixed-size resizing used by classifiers.
  • Because prompts come from a detector rather than a human, the pipeline can scale to high-throughput digitized collections.
  • The multi-region prompting strategy reduces background inside bounding boxes by about ten percentage points relative to single-box prompting.
  • A semi-automatic annotation tool built on the pipeline lets experts correct poor masks with point prompts and expand the fine-tuning dataset.

Reading between the lines

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

  • Inference: the same detector-prompted SAM scheme should transfer to other pressed-organism collections after fine-tuning on a modest sample, since the hard part is domain-specific detection, not the SAM decoder.
  • Inference: because YOLOv10's training labels were derived by thresholding the prior pipeline's masks, a fair reading of the benchmark is that it measures consistency with that pipeline; an independent human-traced test set would reveal whether 0.94 IoU is a true ceiling or a floor.
  • Inference: an ablation that holds SAM2 fixed and varies the prompt source (YOLO boxes, ground-truth boxes, a single random box) would isolate how much of the classification gain comes from detection quality versus segmentation refinement.
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 / 6 minor

Summary. The paper proposes PlantSAM, a pipeline that patches high-resolution herbarium scans, uses YOLOv10 to detect plant regions, feeds bounding-box prompts to SAM1/SAM2, and recombines the patch-level masks into a full-image foreground mask. The authors fine-tune YOLOv10 on bounding boxes derived from their earlier segmentation masks and fine-tune SAM on a curated subset of their earlier segmentation dataset. They report an IoU of 0.94 and a Dice coefficient of 0.97 for PlantSAM2 on a 333-image test set, compare against their own UNet and PlantSAM1, present a visual-evaluation generalization study on 171 out-of-distribution images, and report classification gains of up to +4.36% accuracy and +4.15% F1-score across five botanical traits using ResNet101. They also introduce a semi-automatic annotation tool for mask refinement.

Significance. If the quantitative claims hold, the pipeline is a practical and scalable foreground-isolation tool for herbarium imaging, with a plausible downstream benefit for trait classification. The release of the plant-region detection and segmentation datasets (refs. [6,7]) is a useful resource, and the semi-automatic refinement tool addresses a real operational bottleneck. However, the significance as stated in the abstract is not established by the evidence in this manuscript: segmentation accuracy is measured against masks from the same pipeline lineage, the out-of-distribution evaluation is subjective and based on very small samples, and no external baselines are compared. The contribution is promising but conditional on stronger evaluation.

major comments (4)
  1. [Sec. 3.2, 3.3, 4.2; Figure 7] The provenance of the segmentation ground truth is the central load-bearing issue. YOLOv10 training boxes are generated by thresholding the masks of [3] (Sec. 3.2); SAM1/SAM2 are fine-tuned on a curated subset of the same segmentation dataset [31] (Sec. 3.3); and the 333-image test set [6] is from the same lineage (Sec. 4.2). The IoU/Dice values in Tables 2 and 3 therefore measure agreement with the authors' prior semi-automatic pipeline rather than independently validated segmentation quality. Figure 7's caption calls the ground-truth mask 'manually annotated', which contradicts Sec. 4.2.2's admission that these masks are 'not perfect representations' of ideal segmentation masks. The authors need to clarify the exact annotation protocol and provide an independent manual evaluation on a held-out sample; without this, the headline 0.94/0.97 numbers cannot be interpreted as external accuracy.
  2. [Sec. 4.2.2, Tables 4-6] The generalization study is based on subjective 'visual evaluation' of masks into 'usable' and 'unusable' categories, with no operationalized criteria, no stated number of annotators, and no inter-annotator reliability. Several categories contain only 3-5 images (e.g., Orange background n=3, Long Leaves n=4, Pink background n=5), so percentages such as 33.33%, 60.00%, and 100.0% are not statistically stable. The authors should either add a quantitative metric on this OOD set, for example a small annotated subset evaluated with IoU/Dice under an agreed protocol, or frame Tables 4-6 as illustrative case studies rather than as evidence of generalization.
  3. [Abstract; Sec. 4.2.1] The 'state-of-the-art segmentation performance' claim in the abstract is not supported by the comparison set. The experiments compare PlantSAM1, PlantSAM2, and the authors' own UNet [3] only; no existing herbarium-segmentation method (e.g., Lee et al. [19], Triki et al. [37], GinJinn2 [22], White et al. [40]) is evaluated on the same test set. In addition, Tables 2 and 3 report no error bars, standard deviations, or significance tests, so per-taxon deltas as small as +0.14 IoU and +0.16 Dice (e.g., Litsea and Castanea rows) are not distinguishable from noise. A state-of-the-art claim requires external baselines and repeated runs.
  4. [Sec. 4.3, Table 8] The downstream classification claim of 'consistent performance improvements' is based on what appears to be a single ResNet101 run per condition, with no confidence intervals, no multiple seeds, and no significance testing. Several reported deltas are small (+0.32 accuracy, +0.43 accuracy, +0.78 accuracy), and the 'segmented cropped' condition changes image resolution and scale along with segmentation, so the gain cannot be attributed solely to background removal. The authors should report mean and standard deviation over at least three runs, provide statistical comparisons, and clarify how cropping interacts with input resolution.
minor comments (6)
  1. [Table 3] In the Castanea row, PlantSAM1's Dice is 0.9706 versus UNet's 0.9772, yet the delta is listed as +0.66; it should be approximately -0.66. Please correct and re-check the delta arithmetic in all tables.
  2. [Sec. 4.2.2, Figure 9] The text says 'See Figure 9 for an example of a usable mask', but the Figure 9 caption describes the image as an unusable mask. Please align the text and the caption.
  3. [Sec. 3.3 and Sec. 4.1] Dataset references are inconsistent: Section 3.3 says the SAM fine-tuning subset comes from [7] and [31], while Section 4.1 says training was conducted on 'the curated dataset of 1,476 herbarium segmentation masks [6]'. Since [6] is later used as the test set and [7] as the YOLO detection dataset, please clarify which dataset was actually used for SAM fine-tuning.
  4. [Sec. 4.2.2] The statement that 'PlantSAM2 reduced unusable masks by over 50% compared to UNet' is not uniformly supported by Table 4: for Blue background, 90.91% to 54.55% is a 40% relative reduction, and for Pink background there is no reduction (40.00% to 40.00%). Please rephrase or provide an aggregate statistic.
  5. [Algorithm 1 and Sec. 3.4] The variable is spelled 'patchs', and the threshold criterion (e.g., width/1024 > 3) deserves a one-sentence justification. Please also specify the overlap and padding strategy used when reassembling patches in the Unpatching step, since boundary artifacts at patch seams could affect IoU.
  6. [Sec. 4.1] Training details mention NVIDIA A100 GPUs but do not report inference time or total compute per image. For a pipeline aimed at large-scale herbarium digitization, a runtime estimate would be useful.

Circularity Check

2 steps flagged · score 6.0 of 10

IoU/Dice ground truth comes from the authors' own prior mask pipeline, the same lineage that supervises both YOLOv10 and SAM2, so the 0.94/0.97 headline partly measures self-agreement with those masks; the classification gains rest on independent manual trait labels.

  1. fitted input called prediction [Section 3.2 (YOLOv10 supervision from [3] masks), Section 3.3 (SAM2 fine-tuning on [31]/[3]), Section 4.1 (training on [6]), Section 4.2.1 (IoU/Dice test set [6]), Eqs. (1)-(2)]
    "To train YOLOv10 for this task, we used the Plant Region Detection Dataset [7], which we constructed specifically for this work based on the segmented herbarium image dataset previously published in [3]. ... we then generated bounding boxes around the plant regions by detecting contiguous non-black pixel areas ... To fine-tune the SAM models, we used a curated subset [7] of the Segmentation dataset [31] published in our previous work [3]. ... The quantitative evaluation was conducted using a separate test dataset of 333 herbarium images [6]."

    The mask standard that defines the task is the output of the authors' prior semi-automatic pipeline [3]: YOLOv10's box labels are produced by thresholding those masks ('detecting contiguous non-black pixel areas'), SAM2 is fine-tuned on a curated subset of the same mask dataset, and the 333-image test set [6] is from the same dataset lineage. Both model components are therefore optimized to reproduce the very masks that later serve as ground truth in Eqs. (1)-(2). The reported IoU of 0.94 and Dice of 0.97 thus measure how faithfully the learned pipeline regenerates the prior pipeline's masks, not agreement with independently annotated plant contours.

  2. self citation load bearing [Section 4.2.2 caveat; ground-truth provenance through references [3], [6], and [31], all published by the present authors' group]
    "It is important to interpret these results cautiously, as the datasets used to compute IoU and Dice scores are not perfect representations of ideal segmentation masks. While they are close to the ground truth, we sometimes observe segmentation traits produced by PlantSAM2 that could be considered improvements over the provided annotations."

    The paper's central premise, that the masks from [3]/[6]/[31] are reliable ground truth, is established solely through citations to datasets and papers authored by the present research group, without independent manual re-annotation or an external benchmark for the IoU/Dice evaluation. The cited prior work is the same semi-automatic pipeline whose outputs the models are trained to reproduce, making the reference standard an internal, self-cited artifact rather than an independent ground truth. The paper's own limitation passage admits these masks 'are not perfect representations' and that PlantSAM2 outputs may constitute 'improvements over the provided annotations,' confirming the reference is imperfect.

full rationale

The headline segmentation claim (IoU 0.94, Dice 0.97) is measured against ground-truth masks whose provenance is the authors' own prior semi-automatic pipeline [3]. The chain: [3] produced herbarium segmentation masks via a semi-automatic pipeline (morphological operations plus deep learning); dataset [31] supplies SAM2's fine-tuning masks (1,476 curated images), with [6] used for training in Section 4.1; dataset [7] supplies YOLOv10's supervision by thresholding those same masks into boxes; and the 333-image test set [6] comes from the same dataset lineage. Both model components are therefore optimized to reproduce the masks that then serve as the IoU/Dice reference. The resulting 0.94/0.97 is partly a self-consistency score with the earlier pipeline (the fitted-input-called-prediction pattern), and the paper's Section 4.2.2 caveat ('not perfect representations of ideal segmentation masks') confirms that the reference standard is imperfect and internal. This makes the absolute 'state-of-the-art' number partially circular, with the load-bearing ground truth resting on self-citation. The circularity is only partial, because three parts of the paper rest on independent evidence. (i) The relative ranking PlantSAM2 > UNet > PlantSAM1 is still informative, as all models share the same (flawed) reference. (ii) The OOD generalization evaluation (Tables 4-6) relies on human visual judgment of 'usable' masks rather than self-referential IoU. (iii) The classification experiment uses a separately, manually trait-labeled dataset with a standard ResNet101 baseline, so the accuracy and F1 gains are genuine external evidence that foreground isolation helps trait classification; they do not reduce to the same construction. If the authors had validated IoU/Dice against independent manual re-annotation, the headline would be externally grounded; as written, it largely certifies agreement with their earlier masks.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The load-bearing data, training and evaluation masks and YOLO annotations, all descend from the authors' own earlier segmentation pipeline, so the ledger is dominated by assumptions about the quality and bias of that data. No new physical or conceptual entities are introduced; the pipeline and annotation tool are software artifacts.

free parameters (5)
  • fine-tuning epochs = 80 (SAM), 100 (UNet), 250 (YOLOv10)
    Chosen without sensitivity analysis in Section 4.1; likely not load-bearing but affects performance.
  • patch sizes = 1024, 512, or 256 depending on image width
    Hand-selected threshold in Algorithm 1; impacts context versus resolution trade-off.
  • multi-region vs single-box strategy = multi-region
    Selected after comparing 19 images in Section 3.3; small sample size and subjective comparison.
  • morphological operation kernel = unspecified
    Erosion and dilation are applied to images and masks (Sections 3.1 and 3.3); kernel sizes are not reported, which affects boundary precision.
  • curated training subset size = 1,476 images
    Selected from the full dataset by visual inspection (Section 3.3), introducing selection bias.
assumptions (4)
  • domain assumption The ground-truth masks from the semi-automatic pipeline of Ariouat et al. [3] accurately represent plant regions.
    Used as ground truth for training YOLOv10 (Section 3.2), fine-tuning SAM2 (Section 4.1), and for evaluation (Section 4.2); the paper itself admits the masks are 'not perfect representations'.
  • domain assumption YOLOv10 bounding boxes are sufficient prompts for SAM to produce accurate masks.
    Central to the pipeline design (Section 3.3); point prompts were tried and abandoned because there was no automated placement, not because boxes are optimal.
  • domain assumption Fine-tuning on 9 genera and 2 families generalizes to the test taxa.
    The curated subset is limited to 9 genera and 2 families (Section 3.3), yet the test set includes 11 taxa or groups (Tables 2 and 3).
  • domain assumption The performance gain in classification comes from background removal rather than from the resolution change caused by cropping.
    Section 4.3 compares raw, segmented, and cropped images; cropping changes resolution and aspect ratio, a confound that is not isolated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PlantSAM: An Object Detection-Driven Segmentation Pipeline for Herbarium Specimens." pith.science (2026). https://pith.science/paper/P76D4AO6

@misc{pith2026250716506,
  author       = {Pith},
  title        = {Pith review of: PlantSAM: An Object Detection-Driven Segmentation Pipeline for Herbarium Specimens},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P76D4AO6}},
  note         = {Machine review of arXiv:2507.16506}
}
read the original abstract

Deep learning-based classification of herbarium images is hampered by background heterogeneity, which introduces noise and artifacts that can potentially mislead models and reduce classification accuracy. Addressing these background-related challenges is critical to improving model performance. We introduce PlantSAM, an automated segmentation pipeline that integrates YOLOv10 for plant region detection and the Segment Anything Model (SAM2) for segmentation. YOLOv10 generates bounding box prompts to guide SAM2, enhancing segmentation accuracy. Both models were fine-tuned on herbarium images and evaluated using Intersection over Union (IoU) and Dice coefficient metrics. PlantSAM achieved state-of-the-art segmentation performance, with an IoU of 0.94 and a Dice coefficient of 0.97. Incorporating segmented images into classification models led to consistent performance improvements across five tested botanical traits, with accuracy gains of up to 4.36% and F1-score improvements of 4.15%. Our findings highlight the importance of background removal in herbarium image analysis, as it significantly enhances classification accuracy by allowing models to focus more effectively on the foreground plant structures.

Figures

Figures reproduced from arXiv: 2507.16506 by the authors.

Figure 1
Figure 1. Examples illustrating the diversity in paper color, paper quality and the non-plant elements present in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Plant region detection and segmentation pipeline for herbarium images. Each image is divided into patches, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Single-Box Strategy: A single bounding box [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Comparison of segmentation between the multi-region strategy and the single-box strategy, based on the ratio [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Heat maps of the plant taxa, generated by combining and normalizing their masks. These visualizations [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Illustration of the bounding box pipeline, showing the progression from the original image to the ground [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Examples of herbarium images with various background types, illustrating the diversity in texture and color [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Example of an unusable mask resulting from challenging conditions. The image on the left shows the original [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Illustration of the segmentation-based cropping process. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Example of improving segmentation using a point prompt in the developed application. The user indicates [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 37 canonical work pages

  1. [3]

    Enhancing plant morphological trait identification in herbarium collections through deep learning–based segmentation.Applications in Plant Sciences, 13(2):e70000, 2025

    Hanane Ariouat, Youcef Sklab, Edi Prifti, Jean-Daniel Zucker, and Eric Chenin. Enhancing plant morphological trait identification in herbarium collections through deep learning–based segmentation.Applications in Plant Sciences, 13(2):e70000, 2025

    work page 2025

  2. [31]

    Herbarium image segmentation dataset with plant masks for enhanced morphological trait analysis

    Youcef Sklab, Hanane Sklab, Edi Prifti, Eric Chenin, and Jean-Daniel Zucker. Herbarium image segmentation dataset with plant masks for enhanced morphological trait analysis. Dataset, 2024

    work page 2024

  3. [6]

    Digitised herbarium image segmenta- tion dataset.https://doi.org/10.6084/m9.figshare.29538065.v2, 2025

    Florian Castanet, Hanane Ariouat Sklab, Eric Chenin, and Youcef Sklab. Digitised herbarium image segmenta- tion dataset.https://doi.org/10.6084/m9.figshare.29538065.v2, 2025. Dataset

    work page doi:10.6084/m9.figshare.29538065.v2 2025

  4. [19]

    S. Lee, H. Moon, S. Kim, and J. Lee. Multiple kernel-enhanced encoder for effective herbarium image segmen- tation.Electronics Letters, 61:e70155, 2025

    work page 2025

  5. [37]

    Deep learning based approach for digitized herbarium specimen segmentation.Multimedia Tools and Applications, 81, 08 2022

    Abdelaziz Triki, Bassem Bouaziz, Walid Mahdi, Hamdi Hamed, and Jitendra Gaikwad. Deep learning based approach for digitized herbarium specimen segmentation.Multimedia Tools and Applications, 81, 08 2022

    work page 2022

  6. [22]

    Ott and U

    T. Ott and U. Lautenschlager. Ginjinn2: Object detection and segmentation for ecology and evolution.Methods in Ecology and Evolution, 13:603–610, 2022

    work page 2022

  7. [40]

    A. E. White, R. B. Dikow, M. Baugh, A. Jenkins, and P. B. Frandsen. Generating segmentation masks of herbarium specimens and a data set for training segmentation models using deep learning.Applications in Plant Sciences, 8(6):e11352, 2020. Published 2020 Jul 1

    work page 2020

  8. [1]

    A deep learning-based approach for detecting plant organs from digitized herbar- ium specimen images.Ecological Informatics, 69:101590, 2022

    Abdelaziz, Bassem, and Walid. A deep learning-based approach for detecting plant organs from digitized herbar- ium specimen images.Ecological Informatics, 69:101590, 2022

    work page 2022

Show all 47 references
  1. [2]

    Extracting masks from herbarium specimen images based on object detection and image segmentation techniques.Biodiversity Information Science and Standards, 7, 2023

    Hanane Ariouat, Youcef Sklab, Marc Pignal, Régine Vignes Lebbe, Jean-Daniel Zucker, Edi Prifti, and Eric Chenin. Extracting masks from herbarium specimen images based on object detection and image segmentation techniques.Biodiversity Information Science and Standards, 7, 2023

    2023

  2. [4]

    Sukhorukov, Alain Vanderpoorten, and Farid Jabbour

    Guillaume Besnard, Maël Gaudeul, Sylvain Lavergne, Serge Muller, Grégory Rouhan, Andrey P. Sukhorukov, Alain Vanderpoorten, and Farid Jabbour. Herbarium-based science in the twenty-first century.Botany Letters, 165, 2018

    2018

  3. [5]

    A deep learning based approach for automated plant disease classification using vision transformer.Scientific Reports, 12, 2022

    Yaser Borhani, Javad Khoramdel, and Ebrahim Najafi. A deep learning based approach for automated plant disease classification using vision transformer.Scientific Reports, 12, 2022

    2022

  4. [7]

    Plant region detection in digitised herbarium specimens

    Florian Castanet, Hanane Ariouat Sklab, Eric Chenin, and Youcef Sklab. Plant region detection in digitised herbarium specimens. Dataset, 2025

    2025

  5. [8]

    Robustsam: Segment anything robustly on degraded images.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4081–4091, 2024

    Wei-Ting Chen, Yu-Jiet V ong, Sy-Yen Kuo, Sizhuo Ma, and Jian Wang. Robustsam: Segment anything robustly on degraded images.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4081–4091, 2024

    2024

  6. [9]

    A survey of deep convolutional neural networks applied for prediction of plant leaf diseases

    Vijaypal Singh Dhaka, Sangeeta Vaibhav Meena, Geeta Rani, Deepak Sinwar, Kavita, Muhammad Fazal Ijaz, and Marcin Wo´ zniak. A survey of deep convolutional neural networks applied for prediction of plant leaf diseases. Sensors, 21(14):4749, 2021

    2021

  7. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  8. [11]

    A segmentation-guided deep learning framework for leaf counting.Frontiers in Plant Science, 13, 2022

    Xijian Fan, Rui Zhou, Tardi Tjahjadi, Sruti Das Choudhury, and Qiaolin Ye. A segmentation-guided deep learning framework for leaf counting.Frontiers in Plant Science, 13, 2022

    2022

  9. [12]

    Doctoral thesis, University of Turin and Muséum National d’Histoire Naturelle of Paris, 2024

    Gianluca Grasso.Soils associated with herbarium plants: a resource to address the temporal evolution of plant- associated microbiomes. Doctoral thesis, University of Turin and Muséum National d’Histoire Naturelle of Paris, 2024. 16 APREPRINT

    2024

  10. [13]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    2016

  11. [14]

    Semantic segmentation of herbarium specimens using deep learning techniques

    Burhan Rashid Hussein, Owais Ahmed Malik, Wee-Hong Ong, and Johan Willem Frederik Slik. Semantic segmentation of herbarium specimens using deep learning techniques. In Rayner Alfred, Yuto Lim, Haviluddin Haviluddin, and Chin Kim On, editors,Computational Science and Technology...

  12. [15]

    Convolutional neural networks for image-based high-throughput plant phenotyp- ing: a review.Plant Phenomics, 2020:1–15, 2020

    Yufei Jiang and Chenghai Li. Convolutional neural networks for image-based high-throughput plant phenotyp- ing: a review.Plant Phenomics, 2020:1–15, 2020

    2020

  13. [16]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023

    2023 arXiv

  14. [17]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors,Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012

    2012

  15. [18]

    Lecun, L

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  16. [20]

    Roles of natural history collections.Annals of the Missouri Botanical Garden, 83(4):536–545, 1996

    Lewis Meredith. Roles of natural history collections.Annals of the Missouri Botanical Garden, 83(4):536–545, 1996

    1996

  17. [21]

    Computer vision-based phenotyping for improvement of plant productivity: a machine learning perspective

    Kei Mochida, Shunsuke Koda, Koji Inoue, Takuya Hirayama, Seiji Tanaka, Ryosuke Nishii, and Fabio Melgani. Computer vision-based phenotyping for improvement of plant productivity: a machine learning perspective. GigaScience, 8, 2018

    2018

  18. [23]

    T. Ott, C. Palm, R. V ogt, and C. Oberprieler. Ginjinn: An object-detection pipeline for automated feature extraction from herbarium specimens.Applications in Plant Sciences, 8(6):e11351, 2020

    2020

  19. [24]

    Peter H. Raven. Saving plants, saving ourselves.Plants, People, Planet, 1:8–13, 2019

    2019

  20. [25]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenh...

    2024 arXiv

  21. [26]

    Leveraging multimodal- ity for biodiversity data: Exploring joint representations of species descriptions and specimen images using CLIP

    Maya Sahraoui, Youcef Sklab, Marc Pignal, Régine Vignes Lebbe, and Vincent Guigue. Leveraging multimodal- ity for biodiversity data: Exploring joint representations of species descriptions and specimen images using CLIP. Biodiversity Information Science and Standards, 7:e112666, 2023

    2023

  22. [27]

    Very deep convolutional networks for large-scale image recognition

    K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. pages 1–14. Computational and Biological Learning Society, 2015

    2015

  23. [28]

    Towards a deep learning-powered herbarium image analysis platform.Biodiversity Information Science and Standards, 2024

    Youcef Sklab, Hanane Ariouat, Youssef Boudjydah, Yassine Qacami, Edi Prifti, Jean-Daniel Zucker, Régine Vi- gnes Lebbe, and Eric Chenin. Towards a deep learning-powered herbarium image analysis platform.Biodiversity Information Science and Standards, 2024

    2024

  24. [29]

    Sim-net: A multimodal fusion network using inferred 3d object shape point clouds from rgb images for 2d classification.IET Computer Vision, 19(1):e70036, 2025

    Youcef Sklab, Hanane Ariouat, Eric Chenin, Edi Prifti, and Jean-Daniel Zucker. Sim-net: A multimodal fusion network using inferred 3d object shape point clouds from rgb images for 2d classification.IET Computer Vision, 19(1):e70036, 2025

    2025

  25. [30]

    Identification of non-plant elements in herbarium images using yolo

    Youcef Sklab, Hanane Ariouat, Edi Prifti, Eric Chenin, and Jean-Daniel Zucker. Identification of non-plant elements in herbarium images using yolo. InProceedings of the Conférence Africaine sur la Recherche en Informatique et en Mathématiques (CARI), 2024

    2024

  26. [32]

    Digitization of herbaria enables novel research.American Journal of Botany, 104:1281–1284, 2017

    Pamela Soltis. Digitization of herbaria enables novel research.American Journal of Botany, 104:1281–1284, 2017

    2017

  27. [33]

    Sweeney, Binil Sajan, Paul J

    Patrick W. Sweeney, Binil Sajan, Paul J. Morris, Yiming Xia, Aimee Jirasek, Radhakrishnan Srinivasan, Christo- pher J. Grassa, and Charles C. Davis. Large-scale digitization of herbarium specimens: Development and usage of an automated, high-throughput conveyor system.Taxon, 2...

    2018

  28. [34]

    Szegedy, Wei Liu, Yangqing Jia, P

    C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabi- novich. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9, Los Alamitos, CA, USA, jun 2015. IEEE Comput...

    2015

  29. [35]

    K. M. Thompson, R. Turnbull, E. Fitzgerald, and J. L. Birch. Identification of herbarium specimen sheet compo- nents from high-resolution images using deep learning.Ecology and Evolution, 13:e10395, 2023

    2023

  30. [36]

    Deep leaf: Mask r-cnn based leaf detection and segmentation from digitized herbarium specimen images.Pattern Recognition Letters, 150:76–83, 2021

    Abdelaziz Triki, Bassem Bouaziz, Jitendra Gaikwad, and Walid Mahdi. Deep leaf: Mask r-cnn based leaf detection and segmentation from digitized herbarium specimen images.Pattern Recognition Letters, 150:76–83, 2021

    2021

  31. [38]

    Yolov10: Real-time end-to-end object detection.arXiv preprint arXiv:2405.14458, 2024

    Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection.arXiv preprint arXiv:2405.14458, 2024

    2024 arXiv

  32. [39]

    Weaver and Stephen A

    William N. Weaver and Stephen A. Smith. From leaves to labels: Building modular machine learning networks for rapid herbarium specimen analysis with leafmachine2.Applications in Plant Sciences, 11(5):e11548, 2023

    2023

  33. [41]

    Wilde, Jason G

    Brendan C. Wilde, Jason G. Bragg, and William Cornwell. Analyzing trait-climate relationships within and among taxa using machine learning and herbarium specimens.American Journal of Botany, 110(5):e16167, 2023

    2023

  34. [42]

    Sams : One-shot learning for the segment anything model using similar images.2024 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2024

    Fan Yin, Jun Li, Yifei Wei, Wei Zhang, and Chaoyi Xu. Sams : One-shot learning for the segment anything model using similar images.2024 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2024

    2024

  35. [43]

    De- tection and annotation of plant organs from digitised herbarium scans using deep learning.Biodiversity Data Journal, 8:1–10, 2020

    Sohaib Younis, Marco Schmidt, Christian Weiland, Stefan Dressler, Bernhard Seeger, and Thomas Hickler. De- tection and annotation of plant organs from digitised herbarium scans using deep learning.Biodiversity Data Journal, 8:1–10, 2020

    2020

  36. [44]

    Personalize segment anything model with one shot.ArXiv, abs/2305.03048, 2023

    Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot.ArXiv, abs/2305.03048, 2023

    2023 arXiv

  37. [45]

    Deep-learning-based in-field citrus fruit detection and tracking.Horticulture Research, 9, 2022

    Wenli Zhang, Jiaqi Wang, Yuxin Liu, Kaizheng Chen, Huibin Liu, Yulin Dai, Wenbin Wang, Yun Sun, and Wei Guo. Deep-learning-based in-field citrus fruit detection and tracking.Horticulture Research, 9, 2022

    2022

  38. [46]

    Fast segment anything.ArXiv, abs/2306.12156, 2023

    Xu Zhao, Wen-Yan Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything.ArXiv, abs/2306.12156, 2023

    2023 arXiv

  39. [47]

    Unet++: A nested u-net architecture for medical image segmentation.CoRR, abs/1807.10165, 2018

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation.CoRR, abs/1807.10165, 2018. 18

    2018 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.