Pith. sign in

REVIEW 5 major objections 5 minor 74 references

The paper claims that fusing predictions from several fine-tuned detectors into consensus-weighted soft pseudo-labels turns unlabeled industrial waste footage into training data that performs as well as or better than manual annotation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

An ensemble-based soft pseudo-labeling pipeline improves waste detection on the ZeroWaste dataset, beating fully supervised training with the same labeled images.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful applied paper: the pseudo-annotated ZeroWaste-s set and the soft ensemble labeling recipe are new, and the reported gains are consistent—but missing error bars and SSOD baselines keep the headline claim from being bulletproof. the 5 major comments →

arxiv 2508.18799 v2 pith:TV2MFECD submitted 2025-08-26 cs.CV

Robust and Label-Efficient Deep Waste Detection

Hassan Abid , Khan Muhammad , Muhammad Haris Khan This is my paper
classification cs.CV
keywords semi-supervised object detectionpseudo-labelingensemble learningopen-vocabulary object detectionwaste detectionZeroWastelabel efficiencyobject detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 high-quality annotations for unlabeled industrial waste images can be generated by fusing predictions from an ensemble of fine-tuned detectors, and that training on those pseudo-labels matches or beats training on manual labels. On the ZeroWaste dataset, the ensemble soft pseudo-labeling lifts Grounding DINO Swin-B from 51.6 to 54.3 mAP and Swin-T from 45.6 to 49.3 mAP, without any additional human annotation. Detectors trained exclusively on the final pseudo-labeled set outperform the same detectors trained on the manually labeled split: YOLO11 reaches 40.3 versus 34.0 mAP and RT-DETR 39.4 versus 35.1 mAP. The paper also benchmarks open-vocabulary detectors, showing that zero-shot performance is low on cluttered waste images but rises when prompts are iteratively optimized by an LLM. If correct, the work gives a scalable path to annotating large unlabeled recycling-facility video streams.

Core claim

On the paper's own terms, the central discovery is that consensus-weighted soft pseudo-labels from an ensemble of four transformer detectors (Grounding DINO Swin-B, Co-DETR Swin-L, DETA Swin-L, and DINO Swin-L) can serve as an effective substitute for manual annotation in the ZeroWaste industrial waste domain. Predictions from the ensemble are grouped by class, clustered by pairwise IoU, fused with weighted box fusion, and assigned a soft confidence that penalizes spatial spread and rewards agreement of at least two distinct models. The resulting soft confidence is used to weight the student detector's unsupervised loss. The semi-supervised student improves over full supervision by 2.7 mAP o

What carries the argument

The central mechanism is the ensemble soft pseudo-labeling pipeline. Several fine-tuned detectors each predict boxes on an unlabeled image; detections are filtered by confidence, grouped by category, and clustered by pairwise IoU. Clusters containing detections from at least m=2 distinct models are fused via weighted box fusion, giving a fused box and a base confidence. A consensus factor is then computed from the spatial spread of the contributing boxes and the number of agreeing models: s* = s_base * exp(-alpha*spread) * (1 + beta*(|C_models|-2)). Clusters whose soft confidence s* passes a threshold become pseudo-labels, and during student training each pseudo-labeled sample contributes to

Load-bearing premise

The whole pipeline depends on the four fine-tuned detectors failing in different ways, so when two or more of them agree on a box, that box is truly an object rather than a shared mistake.

What would settle it

A direct test is to generate pseudo-labels from the same four detectors but disable the agreement requirement, keeping all high-confidence clusters regardless of how many models contribute; if the mAP advantage over full supervision disappears, consensus weighting is the cause. A sharper test is to manually label a random subset of ZeroWaste-s and compare a detector trained on those manual labels against one trained on an equal-size sample of the ensemble pseudo-labels, measuring mAP on the ZeroWaste-f test set.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The unlabeled ZeroWaste-s split (6,212 frames) can be converted into 33,075 pseudo-annotations across 6,065 images, providing a large training resource without manual annotation effort.
  • Semi-supervised training with ensemble soft pseudo-labels improves detection on all four classes, including the rare metal class, suggesting the approach does not simply reinforce majority-class bias.
  • Detectors outside the ensemble (YOLO11 and RT-DETR) benefit from training on the pseudo-label set, showing that the pseudo-labels transfer across architectures rather than overfitting to the ensemble's own student.
  • The new fully supervised baselines (up to 51.6 mAP, more than double the previous 24.2 mAP best) give the waste-detection community a stronger reference point for future comparisons.
  • The zero-shot OVOD evaluation indicates that even with LLM-optimized prompts, open-vocabulary models remain far behind fine-tuned detectors on cluttered industrial waste, pointing to domain shift as a major bottleneck.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because every ensemble member is fine-tuned on the same labeled split, part of the reported improvement could be self-distillation from a single teacher family rather than genuinely new label information; an independent fold experiment would reveal how much of the gain comes from consensus versus from the soft weighting itself.
  • The consensus criterion could be ported to other dense detection domains where labeling is scarce, such as sorting construction debris or monitoring conveyor streams, whenever several independently trained detectors are available.
  • The soft-confidence weighting may act as a regularizer by down-weighting noisy boxes; a testable extension would compare the method against simply training on a larger, randomly filtered set of hard pseudo-labels of equal size.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper addresses waste detection on the ZeroWaste industrial dataset. It first benchmarks several open-vocabulary detectors in a zero-shot setting, showing that class-only prompts perform poorly and that LLM-optimized prompts improve accuracy. It then fine-tunes modern closed-set detectors, reporting a new strong baseline of 51.6 mAP. The main contribution is a semi-supervised framework that generates soft pseudo-labels on the unlabeled ZeroWaste-s subset by fusing predictions from an ensemble of fine-tuned detectors via weighted box fusion and a consensus-aware confidence reweighting. The authors report that training detectors with these pseudo-labels improves over fully supervised training, both for the same architecture (Table 3) and for other architectures trained exclusively on the final pseudo-label set (Table 6). The paper also releases the generated pseudo-annotations and code.

Significance. If the reported results are robust, the paper would be a practical contribution to label-efficient waste detection, providing updated baselines and a pseudo-annotation pipeline that could reduce manual annotation cost. The systematic OVOD prompt study and the detailed reproducibility appendix (Appendix D) are strengths. However, the central claim that pseudo-labels can surpass manual labels is currently supported only by single-run point estimates, with no statistical significance assessment, no comparison to established semi-supervised object detection methods, and no quantitative validation of pseudo-label quality. These gaps make the contribution premature in its current form, although the underlying ideas are reasonable and the pipeline is clearly described.

major comments (5)
  1. [§3.4, Table 3 and Table 6] The headline gains are presented as single point estimates. For example, Table 3 reports +2.7 mAP for Swin-B and Table 6 reports +6.3/+4.3 mAP for YOLO11/RT-DETR. In detection with small labeled sets, such differences can easily fall within run-to-run variance. Please provide results over at least 3–5 seeds (mean±std) for all key comparisons, and a significance test if possible. This is necessary to support the abstract claim that pseudo-annotations 'surpass fully supervised training.'
  2. [§2, §3.4] The related work cites STAC, Unbiased Teacher, Soft Teacher, and Dense Teacher, but the experiments do not compare against any of these or other semi-supervised object detection methods. Since the paper proposes a semi-supervised framework, positioning it against these established baselines on ZeroWaste-f/s is essential. Without such comparisons, the contribution of the proposed ensemble pseudo-labeling approach over existing SSOD methods is not demonstrated.
  3. [§3.4, Eq. (3), Appendix C] The pseudo-label quality is only shown qualitatively (Fig. 9). The consensus factor in Eq. (3) rewards agreement among at least m models, but this is only meaningful if the ensemble members have sufficiently independent errors. Appendix D shows that all four ensemble members are transformer-based, three share a Swin-L backbone, and all are fine-tuned on the same labeled split, so correlated failure modes are plausible. Please report precision/recall (or at least a precision vs. confidence curve) of P_Ensemble and P_Final on a human-annotated sample of ZeroWaste-s, with per-class breakdowns. This is particularly important for the rare metal class, which has only 394 pseudo-boxes in Table 5; the reported +1.5/+1.9 AP gains for metal in Table 4 could easily be noise.
  4. [Table 6] The result that YOLO11 and RT-DETR trained exclusively on P_Final outperform the same models trained on manual annotations is surprising, since P_Final is generated by a teacher whose mAP on the test set is 51.6. To rule out confounds, report the number of images and boxes in P_Final vs. ZeroWaste-f, the per-class label distribution, and ideally a control experiment where a model is trained on ZeroWaste-f with labels filtered to the same confidence/box distribution. A discussion of why pseudo-labels can provide a better training signal than human labels is needed.
  5. [Algorithm 1, §3.4] Algorithm 1 requires an initial confidence threshold tau, but the implementation details only specify tau_f=0.35, theta=0.65, m=2, alpha=5.0, beta=0.1, p=2.0, lambda=2.0, gamma=1.0. The value of tau is missing, and the relationship between tau, tau_f, and the final threshold 0.4 used for P_Final is unclear. This harms reproducibility.
minor comments (5)
  1. [Algorithm 1, Eq. (2)] The spread formula in Algorithm 1 uses a denominator of K (number of models), but Eq. (2) uses |C| (number of detections in the cluster). The algorithm should be consistent with Eq. (2).
  2. [§3.4] The sentence 'All experiments were conducted using Grounding DINO (Swin-B), which achieved the overall highest performance in the fully supervised setting (Table 2)' is inaccurate because Table 2 shows three models tied at 51.6 mAP. Rephrase to 'one of the top-performing models.'
  3. [Appendix C, Fig. 9] Figure 9 refers to 'ground-truth annotations' on ZeroWaste-s, but these are manually verified annotations created for qualitative comparison. Please clarify the annotation procedure and whether these labels are released.
  4. [Tables 3 and 4] The mAP notation is inconsistent: 'mAP' vs. 'mAP@50' vs. 'mAP@75' and 'AP@[50:95]'. Please use a single convention (e.g., mAP, AP50, AP75) throughout.
  5. [Abstract] The statement that pseudo-annotations 'achieve performance gains that surpass fully supervised training' is only demonstrated for YOLO11 and RT-DETR in Table 6; it should be qualified, especially since Table 3 shows gains over supervised training but not a direct 'surpassing' of supervised training in the sense of training exclusively on pseudo-labels.

Circularity Check

0 steps flagged

No significant circularity: the pseudo-label pipeline is evaluated on a held-out labeled split and no claimed result is defined in terms of its own output.

full rationale

The derivation chain is self-contained and non-circular. The paper's central claims are empirical comparisons: supervised baselines are trained on ZeroWaste-f labels and evaluated on the ZeroWaste-f test split (Table 2); semi-supervised models are trained on mixed labeled data and pseudo-labels generated from unlabeled ZeroWaste-s, then evaluated on the same held-out test split (Tables 3, 4); and the transfer evaluation in Table 6 trains independent detectors exclusively on P_Final and tests on ZeroWaste-f test. No equation in Section 3.4 reduces to a fitted target: Eqs. 1-4 define a soft confidence score from base confidence, spatial spread, and model agreement, with hyperparameters (alpha, beta, tau, m, p, lambda, gamma) set a priori and not back-fit to the test set. The pseudo-labels are not defined in terms of the final evaluation metric. There is no load-bearing self-citation; the cited works are external baselines, WBF, and standard detectors, and the authors introduce no uniqueness theorem or ansatz that is imported from their own prior work. The skeptical concern that the four ensemble members may share correlated failure modes (all transformer-based, all fine-tuned on the same labeled split) is a legitimate risk to robustness and external validity, but it does not make any claimed result true by construction; it is a correctness/evidence concern, not derivation-level circularity. Accordingly, no circular step can be quoted from the paper, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

10 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard supervised detection assumptions plus a set of hand-set SSL hyperparameters. No new physical entities or formal axioms are introduced. The main unverified premise is that four detectors trained on the same labeled split have complementary errors, so consensus-based pseudo-labels are denoised rather than correlated.

free parameters (10)
  • tau_f = 0.35
    Final soft confidence threshold for keeping pseudo-labels; set by hand, no ablation.
  • theta = 0.65
    IoU threshold for clustering detections before weighted box fusion.
  • m = 2
    Minimum number of distinct models required in a cluster to retain a pseudo-label.
  • alpha = 5.0
    Spread penalty in the consensus factor cf; controls sensitivity to spatial disagreement.
  • beta = 0.1
    Reward multiplier for number of contributing models in the consensus factor.
  • p = 2.0
    Exponent that weights pseudo-label loss by soft confidence.
  • lambda = 2.0
    Regression loss weight in the unsupervised loss.
  • gamma = 1.0
    Relative weight of unsupervised versus supervised loss.
  • P_final confidence threshold = 0.4
    Conservative confidence cutoff used to build the final 33,075-box pseudo-annotation set.
  • tau (initial confidence threshold) = not reported
    Algorithm 1 requires this threshold but the implementation details in Section 3.4 do not specify it.
axioms (5)
  • domain assumption ZeroWaste-s and ZeroWaste-f are captured under the same conditions and share the same label distribution.
    The SSL pipeline and the indirect evaluation of P_final depend on pseudo-labels from ZeroWaste-s transferring to the ZeroWaste-f test set.
  • domain assumption Four detectors trained on the same labeled split have sufficiently independent errors for consensus to improve label quality.
    Section 3.4 keeps pseudo-labels only when at least m distinct models agree; if errors are correlated, consensus cannot remove shared bias.
  • domain assumption ZeroWaste-f ground-truth annotations are accurate and complete enough to serve as an evaluation oracle.
    All mAP comparisons and the 'surpass fully supervised' claim treat these boxes as truth.
  • domain assumption COCO-style mAP computed from the official evaluation code is a valid measure of detection quality.
    All benchmark conclusions rest on this metric, which is standard but only a proxy for sorting performance.
  • domain assumption The MMD and t-SNE analysis in Appendix A using YOLO embeddings measures meaningful domain shift.
    Used to explain OVOD zero-shot failures as domain shift, but the proxy is itself model-specific.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust and Label-Efficient Deep Waste Detection." pith.science (2026). https://pith.science/paper/TV2MFECD

@misc{pith2026250818799,
  author       = {Pith},
  title        = {Pith review of: Robust and Label-Efficient Deep Waste Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TV2MFECD}},
  note         = {Machine review of arXiv:2508.18799}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Effective waste sorting is critical for sustainable recycling, yet AI research in this domain continues to lag behind commercial systems due to limited datasets and reliance on legacy object detectors. In this work, we advance AI-driven waste detection by establishing strong baselines and introducing an ensemble-based semi-supervised learning framework. We first benchmark state-of-the-art Open-Vocabulary Object Detection (OVOD) models on the real-world ZeroWaste dataset, demonstrating that while class-only prompts perform poorly, LLM-optimized prompts significantly enhance zero-shot accuracy. Next, to address domain-specific limitations, we fine-tune modern transformer-based detectors, achieving a new baseline of 51.6 mAP. We then propose a soft pseudo-labeling strategy that fuses ensemble predictions using spatial and consensus-aware weighting, enabling robust semi-supervised training. Applied to the unlabeled ZeroWaste-s subset, our pseudo-annotations achieve performance gains that surpass fully supervised training, underscoring the effectiveness of scalable annotation pipelines. Our work contributes to the research community by establishing rigorous baselines, introducing a robust ensemble-based pseudo-labeling pipeline, generating high-quality annotations for the unlabeled ZeroWaste-s subset, and systematically evaluating OVOD models under real-world waste sorting conditions. Our code is available at: https://github.com/h-abid97/robust-waste-detection.

Figures

Figures reproduced from arXiv: 2508.18799 by Hassan Abid, Khan Muhammad, Muhammad Haris Khan.

Figure 1
Figure 1. Figure 1: Visual examples from the ZeroWaste dataset. (a) Sample images from ZeroWaste-f [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Zero-shot detection performance on the ZeroWaste-f test set using class-only [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Iterative prompt optimization pipeline. Class-level prompts are enriched with [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Soft labeling pipeline for ensemble-based pseudo-labeling. Predictions from [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Class-only vs. optimized prompts on the ZeroWaste-f test set, using Grounding [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE visualizations of deep feature embeddings extracted from a YOLO model [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrices of zero-shot OVOD models on the ZeroWaste-f test set. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of object detection using class-only and optimized prompts [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of fused pseudo-labels against ground-truth annotations [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

74 extracted references · 59 canonical work pages

  1. [1]

    AMP Robotics.https://www.amprobotics.com/

    AMP Robotics. AMP Robotics.https://www.amprobotics.com/. Accessed: 2020-05-30

    work page 2020

  2. [2]

    Dina Bashkirova, Olga Russakovsky, and Stella X. Yu. Zerowaste dataset: Towards deformable object segmentation in cluttered scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12725–12734,

    work page

  3. [3]

    Bircanoglu, M

    C. Bircanoglu, M. Atay, F. Beser, O. Genc, and M. A. Kizrak. Recyclenet: Intelligent waste sorting using deep neural networks. InProceedings of the International Con- ference on Innovative Intelligent Systems and Applications (INISTA), pages 1–7, Jul

    work page

  4. [4]

    Materials recovery facility

    Encyclopædia Britannica. Materials recovery facility. Online, n.d. URLhttps:// www.britannica.com/technology/materials-recovery-facility. Accessed: 20-02-2025

    work page 2025

  5. [5]

    B. D. Carolis, F. Ladogana, and N. Macchiarulo. Yolo trashnet: Garbage detection in video streams. InProceedings of the IEEE Conference on Evolving and Adaptive Intelligent Systems (EAIS), pages 1–7, May 2020. doi: 10.1109/EAIS48028.2020.912 2693

    arXiv 2020

  6. [6]

    Sara Casao, Fernando Peña, Alberto Sabater, Rosa Castillón, Darío Suárez, Eduardo Montijano, and Ana C. Murillo. Spectralwaste dataset: Multimodal data for waste sorting automation. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5852–5858, 2024. doi: 10.1109/IROS58592.2024.10801 797

    arXiv 2024

  7. [7]

    MMDetec- tion: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetec- tion: Open mmlab detection toolbox and...

    Pith/arXiv arXiv 1906

  8. [8]

    Yolo-world: Real-time open-vocabulary object detection

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16901–16911, June 2024

    work page 2024

  9. [9]

    Domestic trash dataset.https://github.com/dataclust er-labs/Domestic-Trash-Dataset, 2021

    Datacluster-labs. Domestic trash dataset.https://github.com/dataclust er-labs/Domestic-Trash-Dataset, 2021. Accessed: 2025-03-01

    work page 2021

  10. [10]

    Advancing sustainable materials manage- ment: 2018 fact sheet, 2020

    Environmental Protection Agency (EPA). Advancing sustainable materials manage- ment: 2018 fact sheet, 2020. URLhttps://www.epa.gov/facts-and-fig ures-about-materials-waste-and-recycling/national-overvie w-facts-and-figures-materials. 12ABIDET AL.: ROBUST AND LABEL-EFFICIENT DEEP W ASTE DETECTION

    work page 2018

  11. [11]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge.International Journal of Computer Vision, 88(2): 303–338, June 2010

    work page 2010

  12. [12]

    An intel- ligent waste-sorting and recycling device based on improved efficientnet.Interna- tional Journal of Environmental Research and Public Health, 19(23):15987, 2022

    Zhicheng Feng, Jie Yang, Lifang Chen, Zhichao Chen, and Linhong Li. An intel- ligent waste-sorting and recycling device based on improved efficientnet.Interna- tional Journal of Environmental Research and Public Health, 19(23):15987, 2022. doi: 10.3390/ijerph192315987

    work page doi:10.3390/ijerph192315987 2022

  13. [13]

    A review on automated sorting of source-separated municipal solid waste for recycling.Waste Management, 60:56–74, 2017

    Sathish Paulraj Gundupalli, Subrata Hait, and Atul Thakur. A review on automated sorting of source-separated municipal solid waste for recycling.Waste Management, 60:56–74, 2017. doi: 10.1016/j.wasman.2016.09.015

    doi:10.1016/j.wasman.2016.09.015 2017

  14. [14]

    Lvis: A dataset for large vocabulary instance segmentation

    Agrim Gupta, Piotr Dollár, Ross Girshick, et al. Lvis: A dataset for large vocabulary instance segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    work page 2019

  15. [15]

    M. Haamer. Wade-ai dataset. Available at Wade-ai project page, 2020

    work page 2020

  16. [16]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. InPro- ceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017

    work page 2017

  17. [17]

    Yolo evolution: A comprehensive benchmark and architectural review of yolov12, yolo11, and their previous versions, 2025

    Nidhal Jegham, Chan Young Koh, Marwan Abdelatti, and Abdeltawab Hendawi. Yolo evolution: A comprehensive benchmark and architectural review of yolov12, yolo11, and their previous versions, 2025. URLhttps://arxiv.org/abs/2411.002 01

    work page 2025

  18. [18]

    Le, Yunhsuan Sung, Zhen Li, and Tom Duerig

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V . Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision, 2021. URLhttps://arxiv. org/abs/2102.05918

    Pith/arXiv arXiv 2021

  19. [19]

    World Bank, 2018

    Silpa Kaza, Lisa Yao, Perinaz Bhada-Tata, and Frank Van Woerden.What a Waste 2.0: A Global Snapshot of Solid Waste Management to 2050. World Bank, 2018. URL https://datatopics.worldbank.org/what-a-waste/. Accessed: 2020-05-30

    work page 2050

  20. [20]

    The role of the waste sector in the sustainable development goals and the ipcc assessment reports.Österreichische Wasser- und Abfallwirtschaft, 76:300–307, 2024

    Radka Kopecká, Marlies Hrad, and Marion Huber-Humer. The role of the waste sector in the sustainable development goals and the ipcc assessment reports.Österreichische Wasser- und Abfallwirtschaft, 76:300–307, 2024. doi: 10.1007/s00506-024-01034-7

    work page doi:10.1007/s00506-024-01034-7 2024

  21. [21]

    Maria Koskinopoulou, Fredy Raptopoulos, George Papadopoulos, Nikitas Mavrakis, and Michail Maniadakis. Robotic waste sorting technology: Toward a vision-based categorization system for the industrial robotic separation of recyclable waste.IEEE Robotics & Automation Magazine, 28(2):50–60, 2021

    work page 2021

  22. [22]

    Scale-aware trident networks for object detection

    Yanghao Li, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Scale-aware trident networks for object detection. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 6054–6063, 2019. ABIDET AL.: ROBUST AND LABEL-EFFICIENT DEEP W ASTE DETECTION13

    work page 2019

  23. [23]

    Microsoft coco: Common objects in context.European Conference on Computer Vision (ECCV), 2014

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ra- manan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context.European Conference on Computer Vision (ECCV), 2014

    work page 2014

  24. [24]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InProceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017

    2017

  25. [25]

    W. Lin. Yolo-green: A real-time classification and object detection model optimized for waste management. InProceedings of the IEEE International Conference on Big Data (Big Data), pages 51–57, Dec. 2021. doi: 10.1109/BIGDATA52589.2021.9671821

    arXiv 2021

  26. [26]

    F. Liu, H. Xu, M. Qi, D. Liu, J. Wang, and J. Kong. Depth-wise separable convolution attention module for garbage image classification.Sustainability, 14(5):3099, Mar

    work page

  27. [27]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. InEuropean Conference on Com- puter Vision, pages 38–55. Springer, 2024

    work page 2024

  28. [28]

    W. Liu, H. Ouyang, Q. Liu, S. Cai, C. Wang, J. Xie, and W. Hu. Image recognition for garbage classification based on transfer learning and model fusion.Mathematical Problems in Engineering, 2022:1–12, Aug. 2022. doi: 10.1155/2022/4793555

    work page doi:10.1155/2022/4793555 2022

  29. [29]

    doi: 10.3390/su14053099

    work page doi:10.3390/su14053099

  30. [30]

    Unbiased teacher for semi-supervised object detection

    Yen-Cheng Liu, Chia-Yi Ma, Zijian He, Simon Kuo, and Jia-Bin Huang. Unbiased teacher for semi-supervised object detection. InInternational Conference on Learning Representations (ICLR), 2021

    work page 2021

  31. [31]

    Dense teacher: Dense pseudo-labels for semi- supervised object detection

    Yue Liu, Shunping Wang, et al. Dense teacher: Dense pseudo-labels for semi- supervised object detection. InInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  32. [32]

    Y . Liu, Z. Ge, G. Lv, and S. Wang. Research on automatic garbage detection system based on deep learning and narrowband internet of things.Journal of Physics: Confer- ence Series, 1069, Aug. 2018. doi: 10.1088/1742-6596/1069/1/012032

    work page doi:10.1088/1742-6596/1069/1/012032 2018

  33. [33]

    Recycling image classification

    Anthony Martin. Recycling image classification. Online. URLhttp://web.cecs .pdx.edu/~singh/rcyc-web/index.html. Accessed: 20-02-2025

    work page 2025

  34. [34]

    O. A. Mengistu. Smart trash net: Waste localization and classification. Online, 2017. URLhttps://www.semanticscholar.org/paper/Final-Report-% 3A-Smart-Trash-Net-%3A-Waste-LocalizationAwe-Mengistu/58 1fb0f0405c7f0e60610d88ceaceb9af44d8569. Accessed: 22-02-2025

    work page 2017

  35. [35]

    W. Ma, X. Wang, and J. Yu. A lightweight feature fusion single shot multibox detector for garbage detection.IEEE Access, 8:188577–188586, 2020. doi: 10.1109/ACCESS .2020.3031990

    arXiv 2020

  36. [36]

    Scaling open-vocabulary ob- ject detection.Advances in Neural Information Processing Systems, 36:72983–73007, 2023

    Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary ob- ject detection.Advances in Neural Information Processing Systems, 36:72983–73007, 2023

    work page 2023

  37. [37]

    Mittal, K

    G. Mittal, K. B. Yagnik, M. Garg, and N. C. Krishnan. Garbage in images (gini) dataset. Online, 2016. URLhttps://github.com/spotgarbage/spotgarbage-G INI. Accessed: 22-02-2025

    work page 2016

  38. [38]

    Simple open-vocabulary object detection

    Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weis- senborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. InEuropean conference on computer vision, pages 728–755. Springer, 2022. 14ABIDET AL.: ROBUST AND LABEL-EFFICIENT DEEP W ASTE DETECTION

    work page 2022

  39. [39]

    Nms strikes back, 2022

    Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krähenbühl. Nms strikes back, 2022. URLhttps://arxiv.org/abs/2212.06137

    Pith/arXiv arXiv 2022

  40. [40]

    P. F. Proença and P. Simões. Taco: Trash annotations in context for litter detection

    work page

  41. [41]

    Sustainable development goal 12: Responsible consumption and pro- duction

    United Nations. Sustainable development goal 12: Responsible consumption and pro- duction. Online, n.d. URLhttps://sdgs.un.org/goals/goal12. Accessed: 19-02-2025

    work page 2025

  42. [42]

    You only look once: Unified, real-time object detection, 2016

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2016. URLhttps://arxiv.org/abs/1506 .02640

    work page 2016

  43. [43]

    Faster r-cnn: Towards real- time object detection with region proposal networks, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real- time object detection with region proposal networks, 2016. URLhttps://arxiv. org/abs/1506.01497

    Pith/arXiv arXiv 2016

  44. [44]

    Ai-powered robotics for recycling.Recycling News, 2022

    AMP Robotics. Ai-powered robotics for recycling.Recycling News, 2022

    work page 2022

  45. [45]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Christopher Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInterna- tional Conference on Machine Learning (ICML), 2021

    work page 2021

  46. [46]

    Serezhkin

    A. Serezhkin. Drinking waste classification dataset.https://www.kaggle.com /datasets/arkadiyhacks/drinking-waste-classification, 2020

    work page 2020

  47. [47]

    Objects365: A large-scale, high-quality dataset for object detection.Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), 2019

    Shuai Shao, Zeming Zhao, Bo Li, et al. Objects365: A large-scale, high-quality dataset for object detection.Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), 2019

    work page 2019

  48. [48]

    A simple semi-supervised learning framework for object detection.arXiv preprint, 2020

    Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection.arXiv preprint, 2020. URLhttps://arxiv.org/abs/2005.04757

    Pith/arXiv arXiv 2020

  49. [49]

    Automated waste sorting: object detection using convolutional neural networks.IOP Conference Series: Earth and Environmental Science, 337(1): 012048, 2019

    Aleksei Seredkin et al. Automated waste sorting: object detection using convolutional neural networks.IOP Conference Series: Earth and Environmental Science, 337(1): 012048, 2019

    work page 2019

  50. [50]

    Joao Sousa, Ana Rebelo, and Jaime S. Cardoso. Automation of waste sorting with deep learning. In2019 XV Workshop de Visão Computacional (WVC), pages 43–48. IEEE, 2019

    work page 2019

  51. [51]

    Sun and H

    A. Sun and H. Xiao. Thanosnet: A novel trash classification method using metadata. InProceedings of the IEEE International Conference on Big Data (Big Data), pages 1394–1401, Dec. 2020. doi: 10.1109/BigData50022.2020.9378287

    arXiv 2020

  52. [52]

    Garbage classification algorithm based on improved mobilenetv3.IEEE Access, 2023

    Xueyong Tian, Liwei Shi, Yuanqing Luo, and Xinlong Zhang. Garbage classification algorithm based on improved mobilenetv3.IEEE Access, 2023

    work page 2023

  53. [53]

    Weighted boxes fusion: En- sembling boxes from different object detection models.Image and Vision Computing, 107:104117, March 2021

    Roman Solovyev, Weimin Wang, and Tatiana Gabruseva. Weighted boxes fusion: En- sembling boxes from different object detection models.Image and Vision Computing, 107:104117, March 2021. ISSN 0262-8856. doi: 10.1016/j.imavis.2021.104117. URL http://dx.doi.org/10.1016/j.imavis.2021.104117. ABIDET AL.: ROBUST AND LABEL-EFFICIENT DEEP W ASTE DETECTION15

    work page arXiv 2021

  54. [54]

    Models supported by ultralytics yolo.https://docs.ultralytics .com/models/, 2025

    Ultralytics. Models supported by ultralytics yolo.https://docs.ultralytics .com/models/, 2025. Accessed: 2025-08-15

    work page 2025

  55. [55]

    A. H. V o, L. H. Son, M. T. V o, and T. Le. A novel framework for trash classification using deep transfer learning.IEEE Access, 7:178631–178639, 2019. doi: 10.1109/AC CESS.2019.2959033

    arXiv 2019

  56. [56]

    Waste-Robotics.https://wasterobotic.com/

    Waste-Robotics. Waste-Robotics.https://wasterobotic.com/. Accessed: 2020-05-30

    work page 2020

  57. [57]

    triangle of death

    Maria Triassi, Rita Alfano, Maddalena Illario, Antonio Nardone, Ornella Caporale, and Paolo Montuori. Environmental pollution from illegal waste disposal and health effects: A review on the “triangle of death”.International Journal of Environmental Research and Public Health, 12:1216–1236, 2015. doi: 10.3390/ijerph120201216

    work page doi:10.3390/ijerph120201216 2015

  58. [58]

    End-to-end semi-supervised object detection with soft teacher

    Bowen Xu, Hang Shi, Yutong Wu, Chenyang Li, Zizhao Zhang, et al. End-to-end semi-supervised object detection with soft teacher. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    work page 2021

  59. [59]

    Yang and G

    M. Yang and G. Thung. Classification of trash for recyclability status. Technical Re- port 3, CS229 Project Report, Stanford University, 2016. URLhttp://cs229.st anford.edu/proj2016/report/ThungYang-ClassificationOfTra shForRecyclabilityStatus-report.pdf. Accessed: 20-02-2025

    work page 2016

  60. [60]

    Hierarchical waste detection with weakly supervised segmentation in images from recycling plants.Engineering Applications of Artificial Intelligence, 128: 107542, 2024

    Dmitry Yudin, Nikita Zakharenko, Artem Smetanin, Roman Filonov, Margarita Kichik, Vladislav Kuznetsov, Dmitry Larichev, Evgeny Gudov, Semen Budennyy, and Alek- sandr Panov. Hierarchical waste detection with weakly supervised segmentation in images from recycling plants.Engineering Applications of Artificial Intelligence, 128: 107542, 2024

    work page 2024

  61. [61]

    YOLO-MTG: A lightweight YOLO model for multi-target garbage detection

    Zhongyi Xia, Houkui Zhou, Huimin Hu, Haoji Hu, Guangqun Zhang, Junguo Hu, and Tao He. YOLO-MTG: A lightweight YOLO model for multi-target garbage detection. Signal, Image and Video Processing, pages 1–16, 2024. doi: 10.1007/s11760-024-032 20-2. URLhttps://link.springer.com/article/10.1007/s11760 -024-03220-2

    work page doi:10.1007/s11760-024-032 2024

  62. [62]

    Zen Robotics.https://zenrobotics.com/

    Zen Robotics. Zen Robotics.https://zenrobotics.com/. Accessed: 2020- 05-30

    work page 2020

  63. [63]

    DINO: DETR with improved denoising anchor boxes for end-to-end ob- ject detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung- Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end ob- ject detection. InThe Eleventh International Conference on Learning Representations,

    work page

  64. [64]

    Detrs beat yolos on real-time object detection

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16965–16974, June 2024

    work page 2024

  65. [65]

    Open- vocabulary object detection using captions

    Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open- vocabulary object detection using captions. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 14393–14402, 2021. URLhttps://openaccess.thecvf.com/content/CVPR2021/html/ Zareian_Open-Vocabulary_Object_Detection_Using_Captions_ CVPR_202...

    work page 2021

  66. [66]

    Instant-teaching: An end-to-end semi-supervised object detection framework

    Qiang Zhou, Chaohui Yu, Zhibin Wang, Qi Qian, and Hao Li. Instant-teaching: An end-to-end semi-supervised object detection framework. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4081–4090, 2021. URLhttps://openaccess.thecvf.com/content/CV PR2021/html/Zhou_Instant-Teaching_An_End-to-End_Semi-Sup ervised...

    work page 2021

  67. [67]

    Detecting twenty-thousand classes using image-level supervision, 2022

    Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krähenbühl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision, 2022. URLhttps: //arxiv.org/abs/2201.02605

    Pith/arXiv arXiv 2022

  68. [68]

    Construction waste object detec- tion based on improved yolov5 algorithm.Sensors, 23(4):1987, 2023

    Zhen Zhou, Xiaofeng Jin, Lu Chen, and Yajuan Han. Construction waste object detec- tion based on improved yolov5 algorithm.Sensors, 23(4):1987, 2023

    work page 1987

  69. [69]

    rigid plastic container

    Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 6748–6758, October 2023. ABIDET AL.: ROBUST AND LABEL-EFFICIENT DEEP W ASTE DETECTION17 Appendix We provide extended analysis to support our main findings across four sectio...

    work page 2023

  70. [70]

    P. Zhou, Z. Zhu, X. Xu, X. Liu, B. He, and J. Zhang. Towards the urban future: A novel trash segregation algorithm based on improved yolov4. InProceedings of the IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 1526– 1531, Dec. 2021. doi: 10.1109/ROBIO54168.2021.9739288

    work page arXiv 2021

  71. [2018]

    doi: 10.1109/INISTA.2018.8466276

    work page arXiv 2018

  72. [2020]

    URLhttps://arxiv.org/abs/2003.06975

    Pith/arXiv arXiv 2003

  73. [2022]

    URLhttps://openaccess.thecvf.com/content/CVPR2022/p apers/Bashkirova_ZeroWaste_Dataset_Towards_Deformable_Ob ject_Segmentation_in_Cluttered_Scenes_CVPR_2022_paper.pdf

    work page

  74. [2023]

    URLhttps://openreview.net/forum?id=3mRwyG5one

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.