Pith. sign in

REVIEW 4 major objections 4 minor 40 references

Anomaly Object Segmentation with Vision-Language Models for Steel Scrap Recycling

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Finetuning the CLIP image encoder lifts steel-scrap anomaly segmentation from below 1% to 28.6% pixel-level average precision.

desk verdict Plausible applied VLM paper with large reported gains, but the evaluation rests on an unreleased dataset and a vaguely described L-SAM-based ground truth; worth review with requests for clarity. read the letter →

arxiv 2506.13282 v1 pith:A7WIRXXP submitted 2025-06-16 cs.CV

Daichi Tanaka , Takumi Karasawa , Shu Takenouchi , Rei Kawakami This is my paper
classification cs.CV
keywords anomalydetectionsteelscraprecyclingvision-languagemodelsCLIPfinetuningmulti-scalemechanismpromptlearningsemanticsegmentationindustrialvisualinspection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that finetuning a vision-language model's image encoder, rather than keeping it frozen, is the key to detecting non-steel contaminants in cluttered steel scrap images. The authors report that their CLIP-based framework with learnable text prompts and a multi-scale mechanism reaches 28.6% pixel-level average precision, while frozen-encoder VLM methods stay below 1% and a fully supervised Mask R-CNN reaches 10.29%. The practical stake is automated visual inspection for steel recycling, where anomalies are scarce, domain-specific, and hard even for human eyes. If the result holds, supervised VLM finetuning becomes a viable route for niche industrial anomaly detection.

What carries the argument

The central object is the CLIP vision transformer with a finetuned image encoder, paired with learnable text prompts and a Multi-Scale Mechanism (MSM). MSM applies average pooling with different strides to patch tokens, produces pseudo anomaly maps at coarser scales via a scale path guided by text similarity, and sums them before upsampling to the original resolution; it is inserted at each of the four encoder stages. The training supervision is a multiclass classification loss (class-balanced focal loss plus dice loss) against ground-truth masks, which forces the model to distinguish anomaly types rather than a single anomaly class.

What would settle it

Train the same model on a public steel scrap dataset (or a released subset of this one) with independently human-verified pixel masks, and compare average precision; if AP drops to the level of Mask R-CNN or below, the claimed advantage is an artifact of the L-SAM-generated annotations.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a CLIP image encoder can be successfully finetuned for anomaly segmentation when the data are niche and cluttered, provided the text prompts are also learned and multi-scale patch features are integrated. The framework uses multiclass classification over normal, motor, cable, and spray-can classes as supervision, and combines per-stage similarity maps with the outputs of a Multi-Scale Mechanism. Across three runs it averages 85.4% AUROC, 19.7% AUPRO, 28.6% AP, and 35.9% F1-max, surpassing the tested zero/few-shot VLM baselines and the fully supervised baselines. The authors interpret this as evidence that domain-specific knowledge must be injected into the visual encoder itself, not only into prompts.

Load-bearing premise

The evaluation depends on a private dataset whose ground-truth masks were produced by Language Segment-Anything without any reported manual verification or inter-annotator agreement; if those masks are noisy or the three chosen anomaly classes are unrepresentative, the reported gap may not transfer to real operation.

Editorial extensions

If this is right

  • Supervised finetuning of a CLIP image encoder lifts pixel-level average precision for niche steel-scrap anomalies from below 1% (frozen-encoder baselines) to 28.6%.
  • Multiclass anomaly supervision with normal, motor, cable, and spray-can classes helps separate anomaly types that share visual clutter.
  • The Multi-Scale Mechanism adds a real gain: AP rises from 25.3% without MSM to 28.6% with 3x3 and 9x9 pooling, and a spray can in a test image is detected only when MSM is active.
  • On the reported data, the finetuned VLM also outperforms fully supervised segmenters such as Mask R-CNN, which scores 10.29% AP.

Reading between the lines

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

  • The 28.6% AP is measured against ground truth generated by Language Segment-Anything; a natural next test is to re-evaluate with human-verified masks to see how much of the gap is real versus an artifact of the annotation procedure.
  • Because the method finetunes the image encoder, it likely needs retraining per site or camera setup; an implicit question is how few labeled images are enough to retain the advantage.
  • The framework assumes the anomaly classes are known and nameable; extending it to open-set contaminants would require an out-of-distribution or rejection mechanism.
  • The text-prompt formulation 'object which is not steel but X' may itself bias learning; a prompt-ablation study could separate the contribution of language guidance from the visual encoder finetuning.
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 proposes a supervised CLIP-based framework for pixel-level anomaly segmentation in steel scrap recycling. The method finetunes the CLIP image encoder together with learnable text prompts and a multi-scale mechanism (MSM), using multiclass classification and segmentation losses as supervision. On a private steel scrap dataset, the method reports an AP of 28.6%, versus less than 1% for frozen-encoder VLM baselines (WinCLIP, WinCLIP+, AnomalyCLIP) and 10.29% for Mask R-CNN, and the authors attribute the gains to supervised finetuning of the vision encoder. The paper also presents an ablation of the MSM scale settings and a qualitative comparison.

Significance. If the reported numbers are reliable, the paper provides a useful data point: supervised finetuning of the CLIP image encoder with multi-scale feature aggregation and prompt tuning can substantially outperform both zero/few-shot VLM methods and conventional supervised detectors on a niche, cluttered industrial anomaly segmentation task. The strengths include the real-world application, a plausible training recipe, and an ablation of the multi-scale mechanism. However, the central quantitative claim rests on a private dataset whose ground-truth masks are reportedly generated by the same language-guided segmentation model used to produce the input crops, with no manual verification protocol, agreement measure, or data release. The absence of error bars and the use of the test set for model selection (Table 3) further weaken the quantitative conclusions. The core idea is defensible and the paper is clearly written, but the evaluation evidence needs strengthening.

major comments (4)
  1. [Section 4.1 (Dataset and Metrics)] The ground truth for the anomaly images is generated by Language Segment-Anything with the text prompt 'the scrap' ('generated in a similar manner'), with no manual verification, inter-annotator agreement metric, or release of the annotation protocol. Because the proposed model is trained with these masks via L_local and L_MSM (Section 3) and evaluated against the same masks, the reported AP in Table 2 may reflect systematic agreement with L-SAM's segmentation biases rather than true anomaly boundaries. The frozen-encoder baselines (WinCLIP, AnomalyCLIP) do not receive mask supervision and therefore cannot adapt to L-SAM's biases, making the comparison in Table 2 potentially unfair. The authors should provide a human-annotated evaluation subset, or at minimum report agreement statistics between the L-SAM masks and expert annotations.
  2. [Section 4.2, Table 3] The MSM configuration (3x3 & 9x9) is selected as the final method by comparing all configurations on the same test set used for the main reported results; no validation split is described. This constitutes test-set tuning and can overestimate the performance of the chosen configuration. The authors should either use a separate validation set for model selection or present the results for all configurations without designating one as the proposed method's final setting.
  3. [Section 4.2, Tables 2 and 3] All quantitative results are reported as averages of three runs, but no standard deviations, confidence intervals, or individual run values are given. In the MSM ablation, the differences between configurations are small (e.g., AP 28.6 vs 27.1 for 3x3 & 9x9 vs 7x7 & 13x13); without variance estimates, the claim that 3x3 & 9x9 is the best configuration is not statistically supported. Please report the spread of the three runs or a significance test.
  4. [Section 4.1 (Dataset and Metrics)] The statement that 'frames in which the entire truck was correctly captured were manually selected' removes potential L-SAM crop failures from the evaluation set before the ground-truth masks are generated. This filtering makes the test set easier than a real deployment scenario in which the crop step may fail. The authors should report how many frames were discarded, and ideally evaluate the full pipeline including crop failures, to assess robustness.
minor comments (4)
  1. [Section 4.1 (Anomaly detection model)] The text says 'an input image size of 518' while Figure 1 states that patch images are set to 512×512; please clarify the actual input resolution and correct any typo.
  2. [Table 1] The table formatting is broken: the test-set anomaly breakdown by class is missing, and the row 'CableMotor Spray can' appears outside the table. Please present the per-class counts for both train and test.
  3. [Section 4.2] The text contains a typo: 'comparision' should be 'comparison'. Also, 'did not perform as good' should be 'did not perform as well'.
  4. [Section 3] The word 'Psuedo' in 'Psuedo anomaly maps' should be 'Pseudo'.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the 28.6% AP claim is empirical, not constructed; the main caveat is unverified L-SAM-derived ground truth, which is a validity risk but not circularity by construction.

full rationale

The central claim (Table 2: AP 28.6% vs less than 1.0% for frozen-encoder VLMs) is an empirical measurement on a held-out test split (Section 4.1, Table 1), not a result derived from definitions, so the equation-level circularity checks do not apply. The training objective (class-balanced focal loss and dice loss between text-image similarity maps and ground truth, Section 4.1) and the evaluation metrics are standard supervised procedures; the model output is not defined in terms of the metric, and the ground truth is not defined in terms of the model output. The most important validity caveat is in Section 4.1: 'The ground truth for the anomaly images was generated in a similar manner' — i.e., by Language Segment-Anything with the text prompt 'the scrap' — and no manual verification or inter-annotator agreement protocol is reported. If the L-SAM masks are noisy or biased toward language-guided segmentation, the gap over frozen-encoder methods could be inflated, and the claim would not transfer to true anomaly boundaries. This is a data-quality and benchmark-representativeness concern, not a circular reduction: the labels come from an external tool, all baselines are scored on the same labels, and the proposed model's output does not feed back into label generation. Two minor tuning issues fall below the circularity bar: the MSM scale size (3x3 and 9x9) was selected as the best row of Table 3 on the same test set, a mild selection-on-test-set effect; and reference [21] (sharing author R. Kawakami) is cited only as motivation ('Inspired by these studies, we also employ multi-scale feature extraction'), while MSM itself is attributed to external works [24, 25], so the self-citation is not load-bearing. The comparison against WinCLIP and AnomalyCLIP is deliberately asymmetric since those zero/few-shot models never see the training labels, which the paper acknowledges, but this is a framing choice rather than a circular step. No uniqueness theorem, ansatz-imported-via-citation, or renaming-of-known-result pattern appears. Score 1 reflects the tuning and unverified ground truth without any identified circular step.

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

The paper introduces no new physical entities or mathematical axioms. Its central claim rests on standard deep learning assumptions, on a private steel scrap dataset, and on several hand-set hyperparameters. The MSM scale selection in Table 3 is performed on the same private evaluation setup, so the final 28.6% AP is partly the product of model selection rather than a parameter-free derivation.

free parameters (5)
  • Stage loss weights for MSM = 0.1 for stages 1-3, 0.7 for stage 4
    Hand-set weights for combining MSM losses; no sensitivity analysis or validation-based selection is reported (Section 4.1).
  • MSM scale sizes = 3x3 and 9x9
    Chosen after comparing 5x5 & 11x11 and 7x7 & 13x13 in Table 3, using the same private evaluation setup; this is model selection on the test data.
  • Learnable text prompt shape = length 16, depth 9, embedding length 4
    Hand-set prompt-tuning hyperparameters; no ablation or search is reported (Section 4.1).
  • Feature extraction layers = layers 6, 12, 18, 24
    Chosen for the ViT-L/14@336px encoder with no stated criterion (Section 4.1).
  • Training schedule = 15 epochs, learning rate 1e-6, batch size 4
    Fixed hyperparameters, no search or schedule analysis reported (Section 4.1).
assumptions (5)
  • domain assumption Language Segment-Anything produces segmentations accurate enough to serve as ground truth for training and evaluation.
    The paper uses L-SAM to crop truck beds and to generate ground-truth masks 'in a similar manner' (Section 4.1), without reporting a validation of mask quality.
  • domain assumption The manually selected frames in which the entire truck is correctly captured are representative of real recycling operations.
    Selection was manual and no distribution or difficulty statistics are reported (Section 4.1).
  • domain assumption CLIP's aligned text-image latent space remains a valid supervision signal after finetuning the image encoder while keeping the text encoder frozen.
    The method computes cosine similarity between tuned text embeddings and finetuned image patch features (Section 3), assuming the alignment is preserved.
  • domain assumption A few hundred labeled anomaly images are sufficient to finetune the visual encoder for niche anomaly discrimination.
    The paper states 'larger models are better few-shot learners' (Section 3) but provides no learning-curve or sample-complexity evidence.
  • ad hoc to paper Averaging the three anomaly prompt weights during binary testing preserves multiclass information.
    Section 4.1 states 'The weights for the three anomaly prompts were averaged,' but no analysis shows this equals multiclass detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Anomaly Object Segmentation with Vision-Language Models for Steel Scrap Recycling." pith.science (2026). https://pith.science/paper/A7WIRXXP

@misc{pith2026250613282,
  author       = {Pith},
  title        = {Pith review of: Anomaly Object Segmentation with Vision-Language Models for Steel Scrap Recycling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A7WIRXXP}},
  note         = {Machine review of arXiv:2506.13282}
}
read the original abstract

Recycling steel scrap can reduce carbon dioxide (CO2) emissions from the steel industry. However, a significant challenge in steel scrap recycling is the inclusion of impurities other than steel. To address this issue, we propose vision-language-model-based anomaly detection where a model is finetuned in a supervised manner, enabling it to handle niche objects effectively. This model enables automated detection of anomalies at a fine-grained level within steel scrap. Specifically, we finetune the image encoder, equipped with multi-scale mechanism and text prompts aligned with both normal and anomaly images. The finetuning process trains these modules using a multiclass classification as the supervision.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    Anomaly Object Segmentation with Vision-Language Models for Steel Scrap Recycling

    INTRODUCTION Steel is produced by forming crude steel, which emits a large amount of carbon dioxide (CO2). The steel industry accounts for 7 to 9% of global CO2 emissions (2020) [1]. Green- house gases, including CO2, significantly contribute to global warming, necessitating emission reductions. The recycling of steel scrap has gained attention as a solut...

    work page Pith review arXiv 2020

  2. [2]

    In particular, CLIP [15] is often utilized due to its aligned feature space for text and images

    RELATED WORK Anomaly detection with foundation modelsSince the ad- vent of vision foundation models [15, 16, 17, 20], these mod- els have been explored for anomaly detection. In particular, CLIP [15] is often utilized due to its aligned feature space for text and images. For instance, Jeonget al.[11] and Chenet al.[10] utilize the similarity between text ...

    work page

  3. [3]

    object which is made of steel

    METHOD We have three main challenges: the uniqueness of the data, data scarcity, and the cluttered nature of the images. To address these challenges, we propose an anomaly detection model based on a VLM. VLMs [15, 16, 17], pretrained on hundreds of millions of images, offer the ability to address data scarcity with their vast knowledge base. Larger models...

    work page

  4. [4]

    Setup Dataset and Metrics.The images used in this study were collected from a steel scrap recycling site and provided by anonymized

    EXPERIMENT 4.1. Setup Dataset and Metrics.The images used in this study were collected from a steel scrap recycling site and provided by anonymized. The images were captured using a 4K camera mounted on a crane that transports the steel scrap. Only Table 1. Breakdown of the dataset used for training and test- ing in our experiment. Each number represents ...

    work page

  5. [5]

    CONCLUSION We have presented a framework combining an anomaly de- tection model based on CLIP [15], one of the vision-language foundation models [16, 17] (VLMs) for application in steel scrap recycling. We can detect and localize the anomaly objects in steel scrap more effectively than both finetuned object detection models with segmentation capabilities ...

    work page

  6. [6]

    object which is not steel

    For binary classification, the normal prompt remained the same as in multiclass classification, and the anomaly prompt was set as “object which is not steel.” The loss functions were the class-balanced focal loss [29] and dice loss [30] for classification and segmentation, both adapted for multiclass version. Using these two loss func- tions, we definedL ...

    work page

  7. [7]

    Padim: A patch distribution modeling frame- work for anomaly detection and localization,

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Ro- maric Audigier, “Padim: A patch distribution modeling frame- work for anomaly detection and localization,” inICPR Inter- national Workshops and Challenges, 2021

    work page 2021

  8. [8]

    Climate change and the production of iron and steel,

    World Steel Association, “Climate change and the production of iron and steel,”

    work page

Show all 40 references
  1. [9]

    Scrap and the steel industry,

    Felipe Maciel, “Scrap and the steel industry,” November 2024

    2024

  2. [10]

    A sub- space projection approach to autoencoder-based anomaly de- tection,

    Jinho Choi, Jihong Park, Abhinav Japesh, and Adarsh, “A sub- space projection approach to autoencoder-based anomaly de- tection,”arXiv:2302.07643, 2023

    2023 arXiv

  3. [11]

    Ganomaly: Semi-supervised anomaly detection via adversar- ial training,

    Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon, “Ganomaly: Semi-supervised anomaly detection via adversar- ial training,” inACCV, 2018

    2018

  4. [12]

    Skip-ganomaly: Skip connected and adversari- ally trained encoder-decoder anomaly detection,

    Samet Akçay and Amir Atapour-Abarghouei andToby P. Breckon, “Skip-ganomaly: Skip connected and adversari- ally trained encoder-decoder anomaly detection,” inIJCNN, 2019

    2019

  5. [13]

    Anomalyd- iffusion: Few-shot anomaly image generation with diffusion model,

    Teng Hu, Jiangning Zhang, Ran Yi, Yuzhen Du, Xu Chen, Liang Liu, Yabiao Wang, and Chengjie Wang, “Anomalyd- iffusion: Few-shot anomaly image generation with diffusion model,” inAAAI, 2024

    2024

  6. [14]

    Beyond dents and scratches: Log- ical constraints in unsupervised anomaly detection and local- ization,

    Paul Bergmann, Kilian Batzner, Michael Fauser, David Satt- legger, and Carsten Steger, “Beyond dents and scratches: Log- ical constraints in unsupervised anomaly detection and local- ization,”International Journal of Computer Vision, vol. 130, no. 4, pp. 947–969, 2022

    2022

  7. [15]

    Towards total recall in industrial anomaly detection,

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler, “Towards total recall in industrial anomaly detection,” inCVPR, 2022

    2022

  8. [16]

    Segment any anomaly without training via hybrid prompt regularization,

    Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zong- wei Du, Liang Gao, and Weiming Shen, “Segment any anomaly without training via hybrid prompt regularization,” arXiv:2305.10724

    arXiv

  9. [17]

    A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad,

    Xuhai Chen, Yue Han, and Jiangning Zhang, “A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad,”arXiv:2305.17382, 2023

    2023 arXiv

  10. [18]

    Winclip: Zero- /few-shot anomaly classification and segmentation,

    Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer, “Winclip: Zero- /few-shot anomaly classification and segmentation,” inCVPR, 2023

    2023

  11. [19]

    Then, they are further divided into patch images

    to create cropped images of the truck bed from the origi- nal anomaly images. Then, they are further divided into patch images. An example of a cropped image is shown in Fig. 1

  12. [20]

    Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection,

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger, “Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection,” inCVPR, 2019

    2019

  13. [21]

    Spot-the-difference self-supervised pre- training for anomaly detection and segmentation,

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer, “Spot-the-difference self-supervised pre- training for anomaly detection and segmentation,” inECCV, 2022

    2022

  14. [22]

    Learning transferable visual models from nat- ural language supervision,

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever, “Learning transferable visual models from nat- ural language supervision,” inICML, 2021

    2021

  15. [23]

    BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi, “BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” inICML, 2022

    2022

  16. [24]

    BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi, “BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” inICML, 2023

    2023

  17. [25]

    Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection,

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jim- ing Chen, “Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection,” inICLR, 2023

    2023

  18. [26]

    lang-segment-anything,

    Luca Medeiros, “lang-segment-anything,” 2024

    2024

  19. [27]

    Segment anything,

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick, “Segment anything,” inICCV, 2023

    2023

  20. [28]

    Thickness classifier on steel in heavy melting scrap by deep-learning-based image analysis,

    Ichiro Daigo, Ken Murakami, Keijiro Tajima, and Rei Kawakami, “Thickness classifier on steel in heavy melting scrap by deep-learning-based image analysis,”ISIJ Interna- tional, vol. 63, no. 1, pp. 197–203, 2023

    2023

  21. [29]

    Waveseg- net: An efficient method for scrap steel segmentation utilizing wavelet transform and multiscale focusing,

    Jiakui Zhong, Yunfeng Xu, and Changda Liu, “Waveseg- net: An efficient method for scrap steel segmentation utilizing wavelet transform and multiscale focusing,”Mathematics, vol. 12, no. 9, 2024

    2024

  22. [30]

    Au- tomated scrap steel grading via a hierarchical learning-based framework,

    Qifan Tu, Dawei Li, Qian Xie, Li Dai, and Jun Wang, “Au- tomated scrap steel grading via a hierarchical learning-based framework,”IEEE Transactions on Instrumentation and Mea- surement, vol. 71, pp. 1–13, 2022

    2022

  23. [31]

    Pyramid scene parsing network,

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia, “Pyramid scene parsing network,” in CVPR, 2017

    2017

  24. [32]

    Clipsam: Clip and sam collaboration for zero- shot anomaly segmentation,

    Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen, “Clipsam: Clip and sam collaboration for zero- shot anomaly segmentation,”Neurocomputing, vol. 618, pp. 129122, 2025

    2025

  25. [33]

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

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

    2021

  26. [34]

    The road less scheduled,

    Aaron Defazio, Xingyu Yang, Harsh Mehta, Konstantin Mishchenko, Ahmed Khaled, and Ashok Cutkosky, “The road less scheduled,”arXiv:2405.15682, 2024

    2024 arXiv

  27. [35]

    Decoupled weight decay regularization,

    Ilya Loshchilov and Frank Hutter, “Decoupled weight decay regularization,” inICLR, 2017

    2017

  28. [36]

    Class-balanced loss based on effective number of samples,

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie, “Class-balanced loss based on effective number of samples,” inCVPR, 2019

    2019

  29. [37]

    Dice loss for data-imbalanced NLP tasks,

    Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li, “Dice loss for data-imbalanced NLP tasks,” in ACL, 2020

    2020

  30. [38]

    Yolov8 by ultralytics,

    “Yolov8 by ultralytics,” https://github.com/ultralytics/ultralytics, 2023

    2023

  31. [39]

    Yolov11 by ultralytics,

    “Yolov11 by ultralytics,” https://github.com/ultralytics/ultralytics, 2024

    2024

  32. [40]

    Mask r-cnn,

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Gir- shick, “Mask r-cnn,” inIn Proc. ICCV, 2017

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.