Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

A multi-modal dataset for insect biodiversity with imagery and DNA at the trap and individual level

T0 review · 3 major / 4 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 MassID45 pairs 45 unsorted Malaise-trap samples with bulk images, DNA metabarcoding, individual barcodes, and 17,937 expert-labeled segmentation masks, and shows that fine-tuned instance segmentation models outperform zero-shot detectors…

desk verdict MassID45 is a genuinely useful and novel dataset, honestly documented, but the ground-truth completeness gap means the benchmark numbers should be treated as provisional until per-sample completeness metadata is added. read the letter →

arxiv 2507.06972 v1 pith:7JEYU3YU submitted 2025-07-09 cs.CV

Johanna Orsholm , John Quinto , Hannu Autto , Gaia Banelyte , Nicolas Chazot , Jeremy deWaard , Stephanie deWaard , Arielle Farrell
show 21 more authors
Brendan Furneaux Bess Hardwick Nao Ito Amlan Kar Oula Kalttopää Deirdre Kerdraon Erik Kristensen Jaclyn McKeown Tommi Mononen Ellen Nein Hanna Rogers Tomas Roslin Paula Schmitz Jayme Sones Maija Sujala Amy Thompson Evgeny V. Zakharov Iuliia Zarubiieva Akshita Gupta Scott C. Lowe Graham W. Taylor
This is my paper · ORCID
classification cs.CV
keywords insectbiodiversityMalaisetrapbulksampleinstancesegmentationDNAmetabarcodingtinyobjectdetectionmulti-modaldatasettaxonomicannotation
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 presents MassID45, a dataset of 45 unsorted Malaise-trap arthropod samples from Sweden and Finland, each with a high-resolution bulk image, DNA metabarcoding, and, after sorting, individual images and COI barcodes for 35,586 specimens. The load-bearing contribution is a set of 17,937 segmentation masks with expert taxonomic labels on the bulk images, produced by human annotators using watershed seeds and an AI-assisted annotation tool, with labels constrained by sample-specific DNA-derived taxonomies. The authors argue this resource closes a gap: ecological surveys collect unsorted bulk samples, while most existing training data is at the level of individual pinned specimens. They benchmark instance segmentation on the bulk images and report that fine-tuned supervised models reach roughly 43% mask average precision, nearly double the best zero-shot model, showing that expert annotations are what make automated counting of trap samples work.

What carries the argument

The central object is the MassID45 dataset itself: 49 bulk images of 45 unsorted Malaise-trap samples, paired with DNA metabarcoding, individual COI barcodes, and 17,937 expert segmentation masks with taxonomic labels. Three mechanisms carry the argument. First, sample-specific taxonomies built from individual-level DNA barcodes constrain the annotator's label choices, linking molecular evidence to each image annotation. Second, watershed segmentation generates initial masks, and a web-based, AI-assisted annotation tool provides refined masks that annotators manually correct. Third, evaluation uses 512-pixel-square tiling with 60% overlap and slicing-aided hyper-inference with non-maximum merging to reassemble full-image predictions, with remote-sensing area thresholds defining small, medium, and large objects.

What would settle it

Take a dense sample with more than 1,000 true specimens, fully sort and identify every individual, and compare the true count and taxon list against the published MassID45 masks and labels; if the annotated count falls far below the true count and independent expert re-annotation changes family-level labels for a nontrivial fraction of masks, the benchmark's ground-truth accuracy is not adequate for reliable abundance estimates.

Watch

Extended reading notes

Core claim

MassID45 is claimed to uniquely pair molecular and imaging data at both the unsorted sample level and the individual-specimen level for the same 45 trap samples. The dataset's central assertion is that the bulk-image annotations—segmentation masks for 17,937 visible arthropods, each tagged with the most specific taxonomic label the annotator could support, plus confidence expressed through multiple labels—are reliable enough to serve as a benchmark for instance segmentation of tiny, densely packed objects and as training data for automated abundance estimation. The authors further claim that DNA metabarcoding supplies taxonomic composition while bulk images supply countable instances, so the combination can yield absolute, taxon-specific abundance estimates. Benchmarking supports the annotations' value: after fine-tuning, the best supervised detector achieves about 43% mask average precision averaged over IoU thresholds from 50% to 95%, whereas the strongest zero-shot detector, an open-vocabulary detector coupled to a foundation-model segmenter, reaches about 27% at the IoU-averaged threshold. The authors interpret this gap as evidence that generalist models alone are insufficient for this domain.

Load-bearing premise

The load-bearing premise is that the bulk-image annotations are complete and accurate enough to serve as ground truth, even though the paper reports that for samples with more than roughly 250 arthropods annotators missed a substantial number of specimens and each taxonomic label was made in about 10 seconds with low-confidence assignments allowed.

Editorial extensions

If this is right

  • An automated pipeline can now count and taxonomically label unsorted trap samples from a single bulk image, replacing manual sorting for abundance estimates.
  • Combining image-derived counts with DNA metabarcoding composition yields absolute, taxon-specific abundances, which metabarcoding alone cannot provide.
  • The dataset extends instance segmentation benchmarks to tiny objects: about three-quarters of masks fall below the standard 'small' area threshold, and the paper's tiling and hyper-inference recipe transfers to other small-object tasks.
  • The fine-tuned supervised detectors provide pretrained weights that are a practical starting point for other bulk-imaging protocols after further fine-tuning.
  • Zero-shot foundation models, without fine-tuning, are not adequate for this domain, so specialist annotations remain necessary.

Reading between the lines

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

  • The paper's own completeness analysis implies that any model trained on these masks will underestimate counts for high-density samples; raw model counts should be calibrated with a density-dependent correction rather than treated as absolute.
  • Because the taxonomic labels were constrained by DNA barcodes from the very same samples, the image labels are partly DNA-informed; classifiers trained on them may perform worse on samples where no barcode data is available.
  • A natural extension the authors leave implicit is using the 35,586 individual images and barcodes as weak supervision to recover missing masks in dense samples, potentially closing the completeness gap without new manual annotation.
  • Reporting per-taxon count error rather than only mask average precision would make the benchmark directly interpretable for ecological monitoring, since that is the quantity ecologists ultimately need.
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

3 major / 4 minor

Summary. The paper introduces MassID45, a dataset of 45 Malaise-trap samples with bulk images (49 images), DNA metabarcoding, individual-level images and DNA barcodes, and 17,937 expert-annotated instance segmentation masks with hierarchical taxonomic labels. It documents the collection, imaging, annotation workflow, DNA processing, and presents a benchmark comparing zero-shot (CutLER, Grounding DINO+SAM, Florence-2+SAM, Gemini+SAM) and supervised (Mask R-CNN, Mask2Former, Mask DINO) instance segmentation models under a tiling plus SAHI protocol.

Significance. The dataset is a genuinely useful resource: it is among the first to pair bulk-tray imagery with both metabarcoding and individual-level barcodes for the same samples, and it targets the under-studied regime of tiny, densely packed objects. The benchmark is honestly designed (validation-based hyperparameter selection, SAHI merging, clear metrics) and the code and data are made available. If the annotation quality can be established, this will support both ecology and computer vision research. The main caveat is that the benchmark's ground truth has a documented completeness gap and unvalidated label accuracy, so the quantitative AP numbers should be treated as provisional until those issues are addressed.

major comments (3)
  1. [Section 2.3.4, Figure 3a] The paper's own completeness analysis shows that for samples with more than roughly 250 arthropods, the number of annotated masks is substantially lower than the number of individually isolated specimens; no per-sample recall, error decomposition, or correction is provided. Since the test set contains six bulk images, missing annotations in any of those images directly bias the AP values in Table 4 and can affect the relative ranking of models. The authors should quantify annotation completeness per sample and per train/validation/test split, report recall against the individual-specimen counts, and either re-annotate affected samples, exclude low-completeness samples from the benchmark, or clearly restrict the benchmark claims to samples with high completeness.
  2. [Section 2.3.3, Appendix S1.4, Table 1] Taxonomic label accuracy is never validated. Appendix S1.4 instructs annotators to spend at most about 10 seconds per specimen and explicitly permits low-confidence labels; Table 1 shows that only 25.4% of labelled specimens have high-confidence family labels. No inter-annotator agreement, no comparison against the individual DNA barcodes (available for 35,586 specimens), and no second-expert audit are reported. Because the scientific value of the dataset depends on users trusting both masks and labels, the authors should provide at least a subsample-based validation of label accuracy against the DNA-based identifications or a second expert, and discuss how the sample-specific taxonomy may constrain label choices.
  3. [Section 4.1.4, Appendix S2.3, Table 4] The evaluation protocol is inconsistent with the definition of AP. The paper states that AP is the area under the precision-recall curve, but then selects a single confidence threshold per model on the validation set and uses that threshold to filter test-set predictions before computing AP. This makes the reported AP values threshold-dependent and not comparable to standard COCO AP. The authors should either report AP over the full confidence range or clearly present the fixed-operating-point metrics (e.g., precision, recall, F1) instead of AP.
minor comments (4)
  1. [Table 4] The Grounding DINO row appears to contain an extra '2.1' and the reported AP50:5:95 value is inconsistent with the text (27.1 vs 2.1); please verify the table formatting and the values.
  2. [Table 1] The phrase 'Values are showed' should be 'Values are shown', and the meaning of the numbers in parentheses under '# Taxa' should be clarified (presumably total taxa including low-confidence labels).
  3. [Section 2.3.2, Appendix S1.3] The class abbreviation 'u' is introduced as 'unidentifiable' in Section 2.3.2 but Appendix S1.3 uses 'unknown' for the same class; please standardize the terminology.
  4. [Section 4.1.1] The handling of ground-truth instances that straddle tile boundaries is not stated precisely; a sentence clarifying that evaluation is performed on the full-resolution merged predictions against the original ground-truth masks would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark evaluations use held-out test data, and the DNA-derived taxonomies are an external input to annotation rather than an output of the derivation.

full rationale

This paper presents a dataset and a benchmark, not a mathematical derivation. The load-bearing empirical claims are that the annotations are accurate enough to serve as ground truth and that the benchmark results are meaningful. These are supported by standard dataset-construction practice: supervised models were fine-tuned on MassID45 training tiles and evaluated on a held-out test partition whose bulk images were never used for training or hyperparameter selection, with confidence thresholds chosen on a separate validation set. The ground-truth masks and labels were produced by human annotators with AI-assisted tools and are independent of the models being benchmarked. The sample-specific taxonomies were constructed from individual-level DNA barcodes and used to constrain the annotator's label menu, but this is an external and independent modality, not an output of the benchmark; moreover, the reported instance-segmentation experiments only use the coarse 'arthropod' class masks, not the taxonomic labels. The documented incompleteness in Section 2.3.4 (annotated counts falling below true specimen counts in large samples) is a correctness and usability limitation, not circularity: the annotations are not fitted to the benchmark metrics, nor are they defined in terms of the model predictions. No equation is defined in terms of another in a way that makes a result true by construction. Minor self-citations, such as the TORAS annotation tool (ref. [28]) or LIFEPLAN protocol papers, are references to tools and protocols, not load-bearing evidence that would make a central claim reduce to a self-citation. Therefore the paper exhibits no significant circularity; its weaknesses are matters of annotation quality and external validity, not circular reasoning.

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

The paper introduces no new physical entities. The assumptions are practical domain assumptions about DNA barcode accuracy and human annotation quality. The free parameters are mostly disclosed hyperparameters and preprocessing constants, none of which are fitted to test data.

free parameters (3)
  • Confidence threshold per model
    For each benchmark model, a confidence threshold is selected by maximizing F1 on the validation set. This is a legitimate operating-point selection, not a free parameter used to fit the claim, but it does affect the reported TP/FP pixel comparisons in the supplementary.
  • Tile size and zoom factor = 512 × 512 (zoom factor 2)
    Tile size was chosen by evaluating validation AP across zoom factors 1-8. This is an honest hyperparameter selection on the validation partition, disclosed in Appendix S2.1.
  • Watershed threshold = 200 in 8-bit grayscale
    The initial segmentation masks use a fixed pixel-intensity threshold of 200. This is a manually chosen preprocessing constant and affects the annotation workload, but not the final labels.
assumptions (3)
  • domain assumption DNA barcode identifications from BOLD reference database are accurate enough to construct sample-specific taxonomies.
    Section 2.2.2 relies on BOLD ID engine and NCBI taxonomy to build the taxonomy that constrains expert labels. If the reference database has errors, labels inherit them.
  • domain assumption Annotators following the given instructions produce masks accurate enough for benchmark ground truth.
    The entire benchmark relies on human annotation quality, but no inter-annotator agreement or verification of mask quality is provided beyond the completeness comparison in Section 2.3.4.
  • standard math The COI barcode region is a valid proxy for species identity in this sample set.
    Standard assumption in DNA barcoding studies, introduced in Section 1 and used throughout the individual and bulk analyses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A multi-modal dataset for insect biodiversity with imagery and DNA at the trap and individual level." pith.science (2026). https://pith.science/paper/7JEYU3YU

@misc{pith2026250706972,
  author       = {Pith},
  title        = {Pith review of: A multi-modal dataset for insect biodiversity with imagery and DNA at the trap and individual level},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7JEYU3YU}},
  note         = {Machine review of arXiv:2507.06972}
}
read the original abstract

Insects comprise millions of species, many experiencing severe population declines under environmental and habitat changes. High-throughput approaches are crucial for accelerating our understanding of insect diversity, with DNA barcoding and high-resolution imaging showing strong potential for automatic taxonomic classification. However, most image-based approaches rely on individual specimen data, unlike the unsorted bulk samples collected in large-scale ecological surveys. We present the Mixed Arthropod Sample Segmentation and Identification (MassID45) dataset for training automatic classifiers of bulk insect samples. It uniquely combines molecular and imaging data at both the unsorted sample level and the full set of individual specimens. Human annotators, supported by an AI-assisted tool, performed two tasks on bulk images: creating segmentation masks around each individual arthropod and assigning taxonomic labels to over 17 000 specimens. Combining the taxonomic resolution of DNA barcodes with precise abundance estimates of bulk images holds great potential for rapid, large-scale characterization of insect communities. This dataset pushes the boundaries of tiny object detection and instance segmentation, fostering innovation in both ecological and machine learning research.

Figures

Figures reproduced from arXiv: 2507.06972 by the authors.

Figure 2
Figure 2. (a) Imaging setup used to capture bulk images of the MassID45 dataset, including the positioning of the camera, light cube and ring light sources. (b) A representative image captured using the described imaging setup, with the sides trimmed. Each photo included a QR code unique to the sample (Figure 2b). For four samples weighing more than approximately 10 g, we divided the sample into two sorting trays, resulting i… view at source ↗
Figure 3
Figure 3. (a) For each sample, a comparison was made between the number of arthropods annotated in the bulk images and the number of individual specimens isolated from the corresponding samples, here shown for all taxa (left) and restricted to Insecta and Arachnida (right). The green line represents a linear regression fit with a 95% confidence interval between the two quantities, and the dashed grey line indicates a 1:1 rela… view at source ↗
Figure 4
Figure 4. Differences between annotations for object detection (left), instance segmentation (middle), and semantic segmentation (right) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Distribution of insect mask areas for “small” (<144 pixels), “medium” (≥144 but <1024 pixels), and “large” (≥1024 pixels) insects. Counts are adjusted such that the area of a bar is proportional to the count in that bin. The three images show the median masks for small…
Figure 6
Figure 6. Figure 6: Visual instance segmentation results for one example patch from the MassID45 test set. Predicted masks are compared for (a) the top-performing zero-shot model, Grounding DINO; and (b) – (d) the 3 supervised baselines: Mask2Former, Mask DINO, and Mask R-CNN. Above each …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. BioAutoML-NAS: An End-to-End AutoML Framework for Multimodal Insect Classification via Neural Architecture Search on Large-Scale Biodiversity Data

    cs.CV 2025-10 reject novelty 4.0 of 10

    An AutoML/NAS insect classifier that feeds the target order labels into its metadata encoder, making the reported 96.81% accuracy uninformative.

Reference graph

Works this paper leans on

67 extracted references · 48 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bánki, O. et al. Catalogue of life, doi:10.48580/dgjy9 (2024)

    work page doi:10.48580/dgjy9 2024

  2. [2]

    Stork, N. E. How many species of insects and other terrestrial arthropods are there on Earth? Annu. Rev. Entomol. 63, doi:10.1146/annurev-ento-020117-043348 (2018)

    doi:10.1146/annurev-ento-020117-043348 2018

  3. [3]

    Wagner, D. L. Insect declines in the anthropocene. Annu. Rev. Entomol. 65, 457–480, doi:10.1146/annurev-ento-011019- 025151 (2020)

    work page doi:10.1146/annurev-ento-011019- 2020

  4. [4]

    Cardoso, P. et al. Scientists’ warning to humanity on insect extinctions. Biol. conservation 242, 108426 (2020)

    work page 2020

  5. [5]

    D., Raven, P

    Wheeler, Q. D., Raven, P. H. & Wilson, E. O. Taxonomy: Impediment or expedient? Science 303, 285–285, doi:10.1126/science.303.5656.285 (2004)

    work page doi:10.1126/science.303.5656.285 2004

  6. [6]

    L., Hamilton, A

    Pearson, D. L., Hamilton, A. L. & Erwin, T. L. Recovery plan for the endangered taxonomy profession. BioScience 61, 58–63, doi:10.1525/bio.2011.61.1.11 (2011)

    work page doi:10.1525/bio.2011.61.1.11 2011

  7. [7]

    Hebert, P. D. N., Cywinska, A., Ball, S. L. & deWaard, J. R. Biological identifications through DNA barcodes. Proc. Royal Soc. London. Ser. B: Biol. Sci. 270, 313–321, doi:10.1098/rspb.2002.2218 (2003)

    work page arXiv 2003

  8. [8]

    Gharaee, Z. et al. A step towards worldwide biodiversity assessment: the BIOSCAN-1M insect dataset. In Oh, A. et al. (eds.) Proceedings of the 37th International Conference on Neural Information Processing Systems, vol. 36 of NIPS ’23, 43593–43619 (Curran Associates Inc., Red Hook, NY , USA, 2023)

    work page 2023

Show all 67 references
  1. [9]

    Gharaee, Z. et al. BIOSCAN-5M: A multimodal dataset for insect biodiversity. In Globerson, A. et al. (eds.) Advances in Neural Information Processing Systems, vol. 37, 36285–36313 (Curran Associates, Inc., 2024)

    2024

  2. [10]

    Steinke, D. et al. Towards a taxonomy machine – a training set of 5.6 million arthropod images. bioRxiv doi:10.1101/2024.07.15.600863 (2024)

    2024 doi

  3. [11]

    Schneider, S. et al. Bulk arthropod abundance, biomass and diversity estimation using deep learning for computer vision. Methods Ecol. Evol. 13, 346–357, doi:10.1111/2041-210X.13769 (2022)

    2022 doi

  4. [12]

    & V ogler, A

    Fujisawa, T., Noguerales, V ., Meramveliotakis, E., Papadopoulou, A. & V ogler, A. P. Image-based taxonomic classi- fication of bulk insect biodiversity samples using deep learning and domain adaptation. Syst. Entomol. 48, 387–401, doi:10.1111/syen.12583 (2023)

    2023 doi

  5. [13]

    Riaz, T. et al. ecoPrimers: inference of new DNA barcode markers from whole genome sequence analysis. Nucleic Acids Res. 39, e145–e145, doi:10.1093/nar/gkr732 (2011)

    2011 doi

  6. [14]

    Ratnasingham, S. et al. BOLD v4: A centralized bioinformatics platform for DNA-based biodiversity data. In DeSalle, R. (ed.) DNA Barcoding: Methods and Protocols, 403–441, doi:10.1007/978-1-0716-3581-0_26 (Springer US, New York, NY , 2024). 14

    doi

  7. [15]

    Luo, M., Ji, Y ., Warton, D. & Yu, D. W. Extracting abundance information from DNA-based data.Mol. Ecol. Resour. 23, 174–189, doi:10.1111/1755-0998.13703 (2023)

    2023

  8. [16]

    Badirli, S. et al. Classifying the unknown: Insect identification with deep hierarchical Bayesian learning. Methods Ecol. Evol. 14, 1515 – 1530 (2023)

    2023

  9. [17]

    Gong, Z. et al. CLIBD: Bridging vision and genomics for biodiversity monitoring at scale. In International Conference on Learning Representations (2025)

    2025

  10. [18]

    Tong, K. & Wu, Y . Deep learning-based detection from the perspective of small or tiny objects: A survey. Image Vis. Comput. 123, 104471, doi:10.1016/j.imavis.2022.104471 (2022)

    2022

  11. [19]

    & Puech, W

    Tresson, P., Carval, D., Tixier, P. & Puech, W. Hierarchical classification of very small objects: Application to the detection of arthropod species. IEEE Access 9, 63925–63932, doi:10.1109/ACCESS.2021.3075293 (2021)

    2021

  12. [20]

    Gupta, A. et al. OW-DETR: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9235–9244 (2022)

    2022

  13. [21]

    Kirillov, A. et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 4015–4026 (2023)

    2023

  14. [22]

    Banelyte, G. G. et al. Global Malaise Trap Project and LIFEPLAN Malaise sampling. protocols.io doi:10.17504/protocols.io.kqdg3xkdqg25/v2 (2023)

    2023 doi

  15. [23]

    Hardwick, B. et al. LIFEPLAN: A worldwide biodiversity sampling design. PLOS ONE 19, 1–15, doi:10.1371/journal.pone.0313353 (2025)

    2025 doi

  16. [24]

    deWaard, J. R. et al. LIFEPLAN Malaise sample metabarcoding. protocols.io doi:10.17504/protocols.io.5qpvokn3xl4o/v1 (2024)

    2024 doi

  17. [25]

    Steinke, D. et al. Low-cost, high-volume imaging for entomological digitization. ZooKeys 1206, 315–326, doi:10.3897/zookeys.1206.123670 (2024)

    2024

  18. [26]

    Hebert, P. D. N. et al. A sequel to Sanger: amplicon sequencing that scales. BMC genomics 19, 1–14, doi:10.1186/s12864- 018-4611-3 (2018)

    2018 doi

  19. [27]

    Schoch, C. L. et al. NCBI taxonomy: a comprehensive update on curation, resources and tools. Database 2020, baaa062, doi:10.1093/database/baaa062 (2020)

    2020 doi

  20. [28]

    Kar, A. et al. Toronto annotation suite. https://aidemos.cs.toronto.edu/toras (2021)

    2021

  21. [29]

    Microsoft COCO: Common objects in context

    Lin, T.-Y .et al. Microsoft COCO: Common objects in context. In Fleet, D., Pajdla, T., Schiele, B. & Tuytelaars, T. (eds.) Computer Vision – ECCV 2014, 740–755 (Springer International Publishing, Cham, 2014)

    2014

  22. [30]

    Waqas Zamir, S. et al. iSAID: A large-scale dataset for instance segmentation in aerial images. In IEEE/CVF Computer Vision and Pattern Recognition Workshops (CVPRW)(2019)

    2019

  23. [31]

    Gillies, S. et al. Shapely, doi:10.5281/zenodo.5597138 (2025)

    2025 doi

  24. [32]

    Orsholm, J. et al. Massid45: Mixed arthropod sample segmentation and identification. Zenodo, doi:10.5281/zenodo.15479861

    doi

  25. [33]

    European Nucleotide Archive, http://identifiers.org/ena.embl:PRJEB86111 (2025)

    2025

  26. [34]

    BOLD Systems, doi:https://dx.doi.org/10.5883/DS-LPEPA22

    Lifeplan epa. BOLD Systems, doi:https://dx.doi.org/10.5883/DS-LPEPA22

    doi

  27. [36]

    Liu, S. et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In Leonardis, A. et al. (eds.) Computer Vision – ECCV 2024, 38–55 (Springer Nature Switzerland, Cham, 2025)

    2024

  28. [37]

    Ren, T. et al. Grounded SAM: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 doi:10.48550/arXiv.2401.14159 (2024)

    2024 doi arXiv

  29. [38]

    Xiao, B. et al. Florence-2: Advancing a unified representation for a variety of vision tasks. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4818–4829, doi:10.1109/CVPR52733.2024.00461 (2024)

    2024

  30. [39]

    Ravi, N. et al. SAM 2: Segment anything in images and videos. In The Thirteenth International Conference on Learning Representations (2025). 15

    2025

  31. [40]

    Introducing Gemini 2.0: our new AI model for the agentic era

    Google DeepMind. Introducing Gemini 2.0: our new AI model for the agentic era. https://blog.google/technology/ google-deepmind/google-gemini-ai-update-december-2024/#ceo-message/ (2024). [Online; accessed 23-April-2025]

    2024

  32. [41]

    & Girshick, R

    He, K., Gkioxari, G., Dollár, P. & Girshick, R. Mask R-CNN. In 2017 IEEE International Conference on Computer Vision (ICCV), 2980–2988, doi:10.1109/ICCV .2017.322 (2017)

    2017 doi

  33. [42]

    Li, F. et al. Mask DINO: Towards a unified transformer-based framework for object detection and seg- mentation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 3041–3050, doi:10.1109/CVPR52729.2023.00297 (2023)

    2023

  34. [43]

    G., Kirillov, A

    Cheng, B., Misra, I., Schwing, A. G., Kirillov, A. & Girdhar, R. Masked-attention mask transformer for universal image segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1280–1289, doi:10.1109/CVPR52688.2022.00135 (2022)

    2022

  35. [44]

    O., Özkalayci, B

    Ünel, F. O., Özkalayci, B. O. & Çi˘gla, C. The power of tiling for small object detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 582–591, doi:10.1109/CVPRW.2019.00084 (2019)

    2019

  36. [45]

    Ding, J. et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE Transactions on Pattern Analysis Mach. Intell. 44, 7778–7796, doi:10.1109/TPAMI.2021.3117983 (2022)

    2022

  37. [46]

    C., Altinuc, S

    Akyon, F. C., Altinuc, S. O. & Temizel, A. Slicing aided hyper inference and fine-tuning for small object detection. 2022 IEEE Int. Conf. on Image Process. (ICIP) 966–970, doi:10.1109/ICIP46576.2022.9897990 (2022)

    2022

  38. [47]

    & Nguyen, K

    Nguyen, L. & Nguyen, K. YOSCA: Confidence adjustment for better object detection in aerial images. Vietnam. J. Comput. Sci. 1–20, doi:10.1142/S219688882450026X (2025)

    2025 doi

  39. [48]

    & Wang, F

    Lin, J., Lin, H. & Wang, F. STPM_SAHI: A small-target forest fire detection model based on swin transformer and slicing aided hyper inference. Forests13, doi:10.3390/f13101603 (2022)

    2022 doi

  40. [49]

    Gia, B. T. et al. Enhancing road object detection in fisheye cameras: An effective framework integrating sahi and hybrid inference. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 7227–7235, doi:10.1109/CVPRW63382.2024.00718 (2024)

    2024

  41. [50]

    Persistent monitoring of insect-pests on sticky traps through hierarchical transfer learning and slicing-aided hyper inference

    Fotouhi, F.et al. Persistent monitoring of insect-pests on sticky traps through hierarchical transfer learning and slicing-aided hyper inference. Front. Plant Sci. Volume 15 - 2024, doi:10.3389/fpls.2024.1484587 (2024)

    2024

  42. [51]

    What is non-max merging? — Roboflow Blog

    Kondrackis, L. What is non-max merging? — Roboflow Blog. https://blog.roboflow.com/non-max-merging/ (2024). [Online; accessed 23-April-2025]

    2024

  43. [52]

    Galloway, A. et al. Predicting dreissenid mussel abundance in nearshore waters using underwater imagery and deep learning. Limnol. Oceanogr. Methods 20, 233–248, doi:10.1002/lom3.10483 (2022)

    2022 doi

  44. [53]

    & Girshick, R

    Wu, Y ., Kirillov, A., Massa, F., Lo, W.-Y . & Girshick, R. Detectron2. https://github.com/facebookresearch/detectron2 (2019)

    2019

  45. [54]

    A., Stricker, D

    Shehzadi, T., Hashmi, K. A., Stricker, D. & Afzal, M. Z. Object detection with transformers: A review. arXiv preprint arXiv:2306.04670 doi:10.48550/arXiv.2306.04670 (2023)

    2023 doi arXiv

  46. [55]

    Rekavandi, A. M. et al. Transformers in small object detection: A benchmark and survey of state-of-the-art. arXiv preprint arXiv:2309.04902 doi:10.48550/arXiv.2309.04902 (2023)

    2023 doi arXiv

  47. [56]

    Li, X. et al. Transformer-based visual segmentation: A survey. IEEE Transactions on Pattern Analysis Mach. Intell.46, 10138–10163, doi:10.1109/TPAMI.2024.3434373 (2024)

    2024

  48. [57]

    Reminder of the first paper on transfer learning in neural networks, 1976

    Bozinovski, S. Reminder of the first paper on transfer learning in neural networks, 1976. Informatica (Slovenia) 44 (2020)

    2020

  49. [58]

    & Hutter, F

    Loshchilov, I. & Hutter, F. Decoupled weight decay regularization. In ICLR (2017)

    2017

  50. [59]

    Smith, L. N. & Topin, N. Super-convergence: Very fast training of residual networks using large learning rates. arXiv preprint arXiv:1708.07120 doi:10.48550/arxiv.1708.07120 (2017)

    2017 doi arXiv

  51. [60]

    Tan, M. & Le, Q. EfficientNet: Rethinking model scaling for convolutional neural networks. In Chaudhuri, K. & Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning, vol. 97 of Proceedings of Machine Learning Research, 6105–6114 (PMLR, 20...

    2019

  52. [61]

    u”, “d”, or “e

    For each polygon mask: a. Change entity name (shortcut: n) to one of these short names: Short name Meaning Explanation b bug Any arthropod u unknown Could be an arthropod, but I can’t tell from the image d debris Any debris, including e.g. loose legs, wings, etc. e edge Mask c...

  53. [62]

    For each unmasked bug that is fully contained in the image and not outlined with red colour (Figure S1.8): a

    Look through the image for arthropods missed by the watershed algorithm. For each unmasked bug that is fully contained in the image and not outlined with red colour (Figure S1.8): a. Add a new entity and mask as described in 1c ii-iii. b. Adjust the mask as described in 1b

  54. [63]

    5 Figure S1.7

    When you are done, mark all entities as complete, and click “Submit”. 5 Figure S1.7. An example of two insects grouped within the same mask. Here, ‘Correct part of curve’ (shortcut: s) is used to draw a new edge to exclude one of the insects. The next step would be to create a...

  55. [64]

    Check the mask of the non-expert annotator

  56. [65]

    If it looks ok, go to classification

  57. [66]

    How confident should I be of the classification?

    If there are big mistakes (see examples below), first correct the mask (for how to do this, see below and consult the instructions above for the non-expert annotator), then go to classification. Examples of big mistakes: • a visual characteristic important for classification o...

  58. [67]

    Ichneumonidae

    When you want to choose one label: include a label on a higher taxonomic level. The label on the highest taxonomic level will then be interpreted as ‘high confidence’, while lower levels are interpreted as ‘low confidence’. Example: you think the bug you are looking at belongs...

  59. [68]

    Tortricidae

    When you want to choose multiple labels: choose multiple labels on any taxonomic level. The last common ancestor of all your chosen labels will be interpreted as ‘high confidence’, whether or not you include it as a label. Example: You think this bug belongs either to Lepidopt...

Pith tools

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