Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Decoupled Pipeline with Proposal Reranking and Score Fusion for Positive-Unlabeled Marine Species Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read For positive-unlabeled marine detection under source shift, the authors claim that a decoupled pipeline—frozen class-agnostic proposals, a separate fine-tuned classifier, and geometric score fusion—outperforms end-to-end fine-tuning.

desk verdict Honest competition report with a useful decoupled-pipeline recipe and a clear public/private reversal; the soft spots are single-run numbers and a validation proxy that is partly fitted and possibly overlapping the test set. read the letter →

arxiv 2607.18700 v1 pith:I6MN53OO submitted 2026-07-21 cs.CV cs.AIcs.LG

Robert James Brock , Sebastian Maximilian Krupa , Jason Kahei Tam This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords positive-unlabeledlearningobjectdetectionmarinespeciesclassificationdomainshiftdecoupledscorefusiontiledinferenceDINOv3
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 for positive-unlabeled marine species detection under a strong train–test source shift, the winning recipe is to decouple: a frozen class-agnostic detector supplies proposal boxes, a separately fine-tuned vision-transformer classifier labels the crops, and a weighted geometric fusion of the two confidences sets the final ranking. This pipeline achieved a private-leaderboard mAP of 0.1757, placing 12th of 102 teams, with a variant adding a light validity-reranking head scoring best on the public leaderboard. A central methodological claim is that train-derived validation and detector-only metrics are unreliable under sparse labels, so model selection must lean on source-matched external evaluation and leaderboard feedback. The paper also concludes that aggressive confidence filtering and hard pseudo-label pruning are counterproductive because many apparent false positives are real organisms that are simply unannotated.

What carries the argument

The load-bearing mechanism is weighted geometric fusion of detector and classifier confidence: detector weight 0.30, classifier weight 0.65, plus an optional validity-head weight 0.05. Geometric fusion penalizes candidates that score poorly in any stage, keeping proposals that are both detectable and classifiable near the top. Supporting machinery includes the frozen class-agnostic Megalodon proposal stack (tiled plus global passes merged by non-maximum suppression with 2-pixel tile-edge filtering), a LoRA-finetuned DINOv3 ViT-H classifier on 2x-expanded crops, and, for the public-best variant, a locally trained reference-conditioned validity head that scores whether a candidate resembles kn

What would settle it

Take the released code and evaluate the frozen-proposal decoupled pipeline against an end-to-end fine-tuned detector on a fresh batch of framegrab images from the same archive that was not used to build the external validation set; if the end-to-end model achieves higher dense mAP under the competition's top-100 metric, the paper's central claim that decoupling wins under source shift is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that decoupling detection from classification outperforms end-to-end training in the FathomNetCLEF 2026 positive-unlabeled setting. A frozen Megalodon YOLOv8x detector, run globally and on tiled crops with tile-edge filtering, generates class-agnostic bounding-box proposals; a LoRA-finetuned DINOv3 ViT-H classifier assigns species labels to expanded proposal crops; and a weighted geometric fusion of detector and classifier confidence produces the final ranking. The system scored 0.1757 private mAP (12/102), while a variant with a reference-conditioned validity head scored 0.1864 public mAP. The paper treats the pipeline's success as evidence that proposal recall,

Load-bearing premise

The paper's model-selection conclusions all rest on the assumption that the externally built validation set—images matched to the hidden test set by URL namespace and annotations mapped through a manually reviewed species crosswalk—faithfully represents the hidden test distribution; the authors themselves note it did not exactly match the final private split.

Editorial extensions

If this is right

  • Under sparse labels and source shift, end-to-end detection/classification plateaus below a decoupled pipeline; decoupling is a viable default for similar positive-unlabeled tasks.
  • Train-derived validation and detector-only metrics can disagree with leaderboard outcomes; external source-matched evaluation should gate model selection.
  • Aggressive confidence thresholds, negative classes, and hard pruning remove unlabeled true positives; preserving long-tail recall and improving ranking is safer.
  • Detector fine-tuning gains measured in detector-only AP may not transfer through the full proposal–classify–rank pipeline; component decisions must be evaluated end-to-end.
  • A reference-conditioned validity score helps public/offline rankings but not the private split; such signals are useful only as a small fusion prior.

Reading between the lines

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

  • The decoupling recipe likely generalizes beyond marine imagery to any positive-unlabeled detection task where the label space is fine-grained and the test distribution shifts, such as medical or ecological imaging with expert-only annotations.
  • The public/private reversal implies that any offline proxy, however domain-matched, has a ceiling; a multi-split private evaluation would separate ranking-signal quality from split noise.
  • A testable extension is to feed the released validity head's reference-conditioned score into a calibrated soft-filtering step (rather than hard pruning) and measure AR@100 alongside mAP, as the paper's own future work suggests.
  • The observed failure of Bio-CLIP versus DINOv3 suggests that for fine-grained marine categories, vision-only self-supervised representations may transfer better than multimodal encoders whose text side was trained on granular scientific names.
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 / 4 minor

Summary. The paper describes DS@GT ARC's system for the FathomNetCLEF 2026 competition, which combines positive-unlabeled object detection and fine-grained marine species classification under strong train/test domain shift. The final private-leaderboard model (Kaggle mAP 0.1757, 12th of 102 teams) is a decoupled pipeline: a frozen Megalodon YOLOv8x detector generates class-agnostic proposals via global and tiled passes merged with NMS and tile-edge filtering; a LoRA-finetuned DINOv3 ViT-H classifier labels expanded proposal crops; and a weighted geometric fusion of detector and classifier confidence produces the final ranking. A variant adding a locally trained TTN-inspired validity head achieved the best public score (0.1864) and best external MBARI proxy score (0.141452) but slightly lower private score (0.1753). The paper's central methodological claim is that, under sparse labels and source shift, decoupling detection from classification, preserving proposal recall, and using an external MBARI proxy for model selection outperform end-to-end fine-tuning, aggressive filtering, and pseudo-label based student-teacher training.

Significance. If the reported results are reliable, the paper provides a useful, reproducible counterexample to end-to-end fine-tuning for positive-unlabeled detection: a frozen class-agnostic proposal generator plus a separately trained classifier and careful score fusion is competitive and robust. The authors also deserve credit for releasing code and for explicitly discussing the limitations of their validation proxy. However, the load-bearing empirical claims rest on the external MBARI proxy, whose construction may overlap the hidden test set, and on comparisons where the fusion-weight sweep is confounded with the addition of the validity head. These issues directly affect the paper's main lessons, so the manuscript needs revision before the claims are fully defensible.

major comments (4)
  1. [§3.3, Tables 5–6, §5.2–5.5] The external_mbari proxy is constructed by selecting FathomNet images whose URLs match the hidden test set's MBARI framegrab namespace (database.fathomnet.org/static/m3/framegrabs/). No disjointness check is reported between external_mbari and the 1,425 hidden test images. Because the hidden test set is itself an MBARI framegrab subset, exact or near-duplicate images may appear in both, meaning model selection and fusion-weight tuning on external_mbari could constitute selection on the test distribution. This would undermine the 'out-of-distribution validation' framing and the conclusions in Tables 5–6 and §§5.3/5.5. Please provide evidence of disjointness (e.g., URL/hash comparison) or re-run analyses after removing any overlapping images.
  2. [§3.6, Table 6, §5.2] The external gain attributed to the TTN-inspired validity head (0.131830 to 0.141452) is measured after re-sweeping all fusion weights on the same external_mbari set, changing the detector/classifier weights from (1.00, 2.00, 0.00) to (0.30, 0.65, 0.05). The validity head is therefore not the only variable; the improvement could come from re-optimizing the original weights. An ablation should either keep the non-validity weights fixed or use nested cross-validation/weight selection for both conditions. In addition, the private difference (0.1757 vs 0.1753) is 0.0004 with no repeated runs reported, so it is within plausible run-to-run noise and cannot support a strong claim of a 'public/private reversal'.
  3. [§3.3, §5.8] The manually reviewed WoRMS-to-CLEF crosswalk is load-bearing for all external_mbari results, but it is not shipped and no accuracy estimate is reported. If the crosswalk misassigns concepts, the proxy labels are noisy, and every conclusion based on external_mbari (frozen over fine-tuned, no aggressive filtering, validity weight 0.05) is correspondingly weakened. The paper should release the crosswalk and provide at least a manual audit or inter-annotator agreement, and it should quantify how crosswalk uncertainty affects the reported mAP differences.
  4. [§4, §5.8] All mAP numbers, including the main leaderboard results, are single runs without error bars, confidence intervals, or significance tests. For decisions where the reported differences are small (e.g., 0.1757 vs 0.1753; 0.1809 vs 0.1864 on public), the absence of repeated runs or a variance estimate makes it difficult to distinguish genuine improvements from stochastic variation. Please report at least a small number of repeated runs for key comparisons, or clearly label borderline differences as non-significant.
minor comments (4)
  1. [§3.2] Typo: 'Non-Maximum Supression' should be 'Non-Maximum Suppression'.
  2. [§3.3, Table 1] The distinction between external_mbari and external_mbari_strict is clear, but it would help to mention in the table caption that all downstream tables use the full view unless otherwise stated.
  3. [§5.2, Table 9] The row 'Validity-only control 0.055051' is informative, but consider stating explicitly that this control uses the validity score as the sole fusion score, to avoid ambiguity.
  4. [References] Some references are incomplete or are web resources without DOIs (e.g., [6], [7], [22]). Please add access dates and, where possible, persistent identifiers.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the validity-head external-MBARI gain is an in-sample selection effect because fusion weights are tuned on the same proxy; the central decoupled-pipeline/private-leaderboard result remains independently grounded.

  1. fitted input called prediction [§3.6 fusion-weight sweep; Tables 6 and 9]
    "The fusion weights are selected by running inference on the external MBARI set to obtain detector, classifier (and TTN validity scores for best public approach), then sweeping over weight combinations and selecting the combination that maximizes mAP."

    The validity weight (0.05) and the reported external-MBARI gain are jointly chosen: the sweep optimizes mAP on external_mbari, and Tables 6 and 9 then report the resulting 0.131830→0.141452 as evidence that 'the validity head improved external MBARI dense mAP by 0.009622'. The external gain is therefore an in-sample optimum, not an out-of-sample prediction; part of the improvement is attributable to weight selection rather than to the validity head. The public-leaderboard gain (0.1809→0.1864) provides independent confirmation, so the circularity is partial rather than total.

full rationale

The paper's central derivation chain is not circular: the final private-best pipeline is a decoupled frozen-Megalodon proposal stack plus DINOv3 classifier plus geometric fusion, and its 0.1757 private-leaderboard score is an independent external outcome. The main design choices are supported by leaderboard transfers and ablations with public and private scores, and no load-bearing result is justified solely by a self-citation; references to prior DS@GT work are contextual, while the Megalodon and TTN citations are external. The one genuine circularity is limited to the validity-head ablation: the fusion weights, including the validity exponent 0.05, are swept to maximize external-MBARI mAP on the same proxy used in Tables 6 and 9, so the reported external gain (0.131830 to 0.141452) is partly a selection artifact. The public-leaderboard improvement gives independent support, and the private-best model excludes the validity head, so this does not undermine the central decoupled-pipeline claim. External-proxy representativeness (URL-namespace matching, manual crosswalk, no disjointness check) is a real validity risk but is not circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

No physical entities, forces, or conserved quantities are postulated; the only introduced artifact is the local validity head, which carries mixed out-of-sample evidence. The paper's free parameters are design choices fitted on the external MBARI proxy, and the central empirical claims rest on the proxy assumption plus single-run leaderboard numbers.

free parameters (7)
  • Fusion weights (detector/classifier/validity) = 0.30 / 0.65 / 0.05 (public-best); detector/classifier-only for private-best
    Selected by sweeping weight combinations on the external MBARI set and picking the mAP-maximizing point (§3.6, §5.2). The same set is then used to report the validity-head improvement in Table 6.
  • Tile overlap = 30%
    Hand-selected (§3.7); affects recall of small organisms split across tiles.
  • Tile edge-filter epsilon = 2 px
    Hand-selected discard threshold for detections touching tile edges (§3.7).
  • Crop expansion factor = 2.0×
    Chosen via context sweep on external MBARI (Table 14: 2.0 beats 2.5); effectively fitted on the proxy.
  • Validity head design = 6-layer transformer, 10 reference crops, weight 0.05
    Architecture and hyperparameters chosen by hand (§3.5); the 0.05 fusion exponent is part of the fitted sweep.
  • LoRA hyperparameters = rank 16, alpha 16, dropout 0.1, lr 2e-4 / 1e-3, warmup 0.05
    Hand-selected (§3.7); not ablated in the paper.
  • Pseudo-label admission threshold (ablations only) = confidence ≥ 0.5
    Used in the Faster R-CNN pseudo-label experiment (Table 3); not part of the final system.
assumptions (5)
  • domain assumption The external MBARI proxy (URL-matched, manually crosswalked) approximates the hidden test distribution well enough for model selection.
    Stated in §3.3; the entire model-selection methodology rests on it. The authors qualify it in §5.8 as imperfect, and the public/private reversal is direct evidence of its limits.
  • domain assumption The manually reviewed WoRMS-to-CLEF concept crosswalk is accurate; unmapped annotations can be dropped without biasing evaluation.
    §3.3. Dropping unmapped annotations makes external_mbari a partial-label evaluation; if dropped concepts are common on some images, the mAP is distorted in an unknown direction.
  • domain assumption Classifier training crops from matched ground-truth boxes are correctly labeled and representative of detector-proposal geometry.
    §3.1: the classifier trains exclusively on matched ground-truth crops but is applied to detector proposals whose geometry, padding, and noise differ.
  • standard math NMS at IoU 0.5 and weighted geometric mean fusion preserve the ranking that dense mAP rewards.
    Standard tools used in §3.2 and §3.6; not derived in the paper.
  • domain assumption The positive-unlabeled failure model — unannotated organisms act as background and are penalized as false positives — is the correct interpretation of the training loss.
    §3.1: 'The unlabeled organisms act as background, which meant the model is penalized every time it correctly detects a real organism but was not annotated.' This motivates the entire decoupling design.
invented entities (1)
  • TTN-inspired local validity head independent evidence
    purpose: Class-conditional validity score that reranks detector/classifier fusion: does a candidate crop resemble known members of its predicted class?
    Introduced in §3.5; unlike Griffin & Corso's frozen TTN it is locally trained on DINOv3 embeddings. It has falsifiable handles: public leaderboard improvement (0.1809→0.1864) and external MBARI improvement, though the private leaderboard slightly degraded (0.1757→0.1753), so the evidence is mixed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupled Pipeline with Proposal Reranking and Score Fusion for Positive-Unlabeled Marine Species Detection." pith.science (2026). https://pith.science/paper/I6MN53OO

@misc{pith2026260718700,
  author       = {Pith},
  title        = {Pith review of: Decoupled Pipeline with Proposal Reranking and Score Fusion for Positive-Unlabeled Marine Species Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6MN53OO}},
  note         = {Machine review of arXiv:2607.18700}
}
read the original abstract

The FathomNetCLEF 2026 competition combines underwater object detection and fine-grained marine species classification under a positive-unlabeled evaluation setting. The provided training labels are sparse, while the hidden test set is out-of-distribution relative to the training imagery, creating both annotation incompleteness and source-shift challenges. We describe DS@GT ARC's multi-stage system developed for this setting while keeping model training restricted to the data provided by the competition. The final private-leaderboard model uses a frozen Megalodon YOLOv8x detector as a class-agnostic proposal generator, combines global and tiled inference with tile-edge filtering, classifies expanded proposal crops with a LoRA-finetuned DINOv3 ViT-H classifier, and ranks predictions using weighted geometric fusion of detector and classifier confidence. This system placed 12th out of 102 teams. A closely related variant added a locally trained TTN-inspired validity head as a light reranking signal, improving public-leaderboard and proxy-evaluation performance but slightly reducing private-leaderboard performance. Across experiments, the strongest lesson was that train-derived validation and detector-only metrics were not reliable enough for model selection. Instead, we used proxy datasets only for validation and comparison, and combined those signals with leaderboard feedback and targeted ablations. These experiments showed that reserving proposal recall, avoiding over-aggressive filtering, and improving downstream ranking were more effective than fine-tuning the detector or directly training on noisy pseudo-labels. Code: https://github.com/dsgt-arc/fathomnetclef-2026.

Figures

Figures reproduced from arXiv: 2607.18700 by the authors.

Figure 1
Figure 1. Distribution of classes in the training set [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Three examples from the training set with their bounding boxes overlaid the Fathomnet 2025 [6] competition passed multiscale images and full-scaled images centered on the region-of-interest (ROI) into ViT, combining object embeddings with context patch embeddings. 3. Methodology Our pipeline is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Pipeline of our best approach robust vision transformer which showed promising results in previous LifeCLEF competitions [10][20], and image resolution flexibility. Swin-L was chosen as an alternative backbone for its hierarchical feature extractor. DINO-DETR was chosen as an alternative framework to Faster R-CNN since it eliminated the need for anchor boxes, thus reducing complexity. We observed that end-to-end det… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 8 linked inside Pith

  1. [1]

    Katija, E

    K. Katija, E. Orenstein, B. Schlining, L. Lundsten, K. Barnard, G. Sainz, O. Boulais, M. Cromwell, E. Butler, B. Woodward, K. C. Bell, Fathomnet: A global image database for enabling artificial intelligence in the ocean, 2022. URL: https://arxiv.org/abs/2109.14646.arXiv:2109.14646

    arXiv 2022

  2. [2]

    URL: https://ocean.si.edu/planet-ocean/ seafloor/just-how-big-ocean, accessed: 2026-04-22

    Smithsonian Ocean, Just how big is the ocean?, n.d. URL: https://ocean.si.edu/planet-ocean/ seafloor/just-how-big-ocean, accessed: 2026-04-22

    2026

  3. [3]

    Chrobak, K

    L. Chrobak, K. Barnard, Overview of FathomNetCLEF 2026: Positive-unlabeled object detection in marine images, in: Working Notes of CLEF 2026 - Conference and Labs of the Evaluation Forum, 2026

    2026

  4. [4]

    Picek, L

    L. Picek, L. Adam, S. Kahl, R. Bossy, L. Chrobak, H. Goëau, K. Papafitsoros, H. Klinck, W.-P. Vellinga, R. Planqué, T. Denton, K. Barnard, C. Nédellec, L. Deléger, M. Courtin, G. Martellucci, I. Moummad, F. Vinatier, P. Bonnet, A. Joly, Overview of LifeCLEF 2026: Ai challenges for biodiversity understanding and ecosystem management, in: International Conf...

    2026

  5. [5]

    Barnard, L

    eor123, HCL-Jevster, HCL-Rantig, K. Barnard, L. Chrobak, picekl, Fathomnet2024, https://kaggle. com/competitions/fathomnet2024, 2024. Kaggle

    2024

  6. [6]

    L. Chrobak, Looking at the bigger picture: Results from fathomnet’s 2025 kag- gle competition, https://www.fathomnet.org/news/looking-at-the-bigger-picture: results-from-fathomnet-2025-kaggle-competition, 2026. Accessed: 2026-04-29

    2025

  7. [7]

    Ultralytics YOLOv8x Object detection model fine- tuned to detect marine objects on FathomNet

    Monterey Bay Aquarium Research Institute (MBARI), Fathomnet megalodon detector, https: //huggingface.co/FathomNet/megalodon, 2024. Ultralytics YOLOv8x Object detection model fine- tuned to detect marine objects on FathomNet

    2024

  8. [8]

    B. A. Griffin, J. J. Corso, The label imitation game: Turing test network for zero-shot pseudo-label pruning, arXiv preprint (2026)

    2026

Show all 26 references
  1. [9]

    DS@GT ARC, FathomNetCLEF 2026 code release, https://github.com/dsgt-arc/fathomnetclef-2026,

    2026

  2. [10]

    Gustineli, A

    M. Gustineli, A. Miyaguchi, A. Cheung, D. Khattak, Tile-based vit inference with visual-cluster priors for zero-shot multi-species plant identification, 2025. URL: https://arxiv.org/abs/2507.06093. arXiv:2507.06093

    2025 arXiv

  3. [11]

    Ozge Unel, B

    F. Ozge Unel, B. O. Ozkalayci, C. Cigla, The power of tiling for small object detection, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019

    2019

  4. [12]

    F. C. Akyon, S. Onur Altinuc, A. Temizel, Slicing aided hyper inference and fine-tuning for small object detection, in: 2022 IEEE International Conference on Image Processing (ICIP), IEEE, 2022, pp. 966–970. URL: http://dx.doi.org/10.1109/ICIP46576.2022.9897990. doi:10.1109/ic...

    2022

  5. [13]

    Sharma, D

    T. Sharma, D. E. Cline, D. Edgington, Making use of unlabeled data: Comparing strategies for marine animal detection in long-tailed datasets using self-supervised and semi-supervised pre- training, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2024

  6. [14]

    Self-supervised learning approach with Multi-Context Environmental Attention Module (MCEAM) and Hierarchical Auxiliary Classification for marine animal recognition

    dhlee-work, GGGuni, Solution to the CVPR’2025 FathomNet-FGVC challenge, https://github.com/ dhlee-work/fathomnet-cvpr2025-ssl, 2025. Self-supervised learning approach with Multi-Context Environmental Attention Module (MCEAM) and Hierarchical Auxiliary Classification for marine...

    2025

  7. [15]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adapta- tion of large language models, 2021. URL: https://arxiv.org/abs/2106.09685. arXiv:2106.09685

    2021 arXiv

  8. [16]

    Siméoni, H

    O. Siméoni, H. V. Vo, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V. Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. Jég...

    2025 arXiv

  9. [17]

    S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. URL: https://arxiv.org/abs/1506.01497.arXiv:1506.01497

    2016 arXiv

  10. [18]

    Zhang, F

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, H.-Y. Shum, Dino: Detr with improved denoising anchor boxes for end-to-end object detection, 2022. URL: https://arxiv.org/abs/2203.03605. arXiv:2203.03605

    2022 arXiv

  11. [19]

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierar- chical vision transformer using shifted windows, 2021. URL: https://arxiv.org/abs/2103.14030. arXiv:2103.14030

    2021 arXiv

  12. [20]

    Miyaguchi, C

    A. Miyaguchi, C. Maruthaiyannan, C. R. Clark, Ds@gt animalclef: Triplet learning over vit manifolds with nearest neighbor classification for animal re-identification, 2025. URL: https: //arxiv.org/abs/2509.12353.arXiv:2509.12353

    2025

  13. [21]

    Kaggle dataset, accessed 2026-02-19

    Vencerlanz09, Sea animals image dataset, https://www.kaggle.com/datasets/vencerlanz09/ sea-animals-image-dataste/data, 2024. Kaggle dataset, accessed 2026-02-19

    2024

  14. [22]

    J. e. a. Gu, BioCLIP 2.5 Huge, 2026. URL: https://huggingface.co/imageomics/bioclip-2.5-vith14

    2026

  15. [23]

    J. e. a. Gu, BioCLIP 2: Emergent properties from scaling hierarchical contrastive learning, in: The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL: https://arxiv.org/abs/2505.23883

    2025

  16. [24]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative adversarial networks, 2014. URL: https://arxiv.org/abs/1406.2661. arXiv:1406.2661

    2014 arXiv

  17. [25]

    URL: http://www

    PACE, Partnership for an Advanced Computing Environment (PACE), 2017. URL: http://www. pace.gatech.edu

    2017

  18. [2026]

    Implementation, configuration files, and experiment scripts

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.