Pith. sign in

REVIEW 2 major objections 6 minor 27 references

MDWD: A Street-Level Dataset for Municipal Solid Waste Detection in Dense Urban Environments

T0 review · 2 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper introduces MDWD, a street-level benchmark of 3,697 images and 11,461 labeled waste instances across five municipal waste streams, and shows that detectors trained on it reach high accuracy (RF-DETR-M mAP50 94.49%, F1 93.56%).

desk verdict Solid, well-documented dataset paper with a real gap to fill; the benchmark numbers are conditional on annotation consistency that is asserted but unmeasured. read the letter →

arxiv 2608.00257 v1 pith:VZAY37FY submitted 2026-07-31 cs.CV

Andrea Filiberto Lucas , Mark Bugeja , Carl James Debono , Dylan Seychell This is my paper
classification cs.CV
keywords municipalsolidwasteobjectdetectionstreet-leveldatasetbenchmarkYOLORF-DETRclassificationinstance-levelannotation
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 introduces the Maltese Domestic Waste Dataset (MDWD), a street-level collection of 3,697 high-resolution images with 11,461 manually annotated waste instances in five categories that mirror a real municipal collection scheme. It claims that no existing waste dataset simultaneously offers street-level imagery, instance-level bounding boxes, and multi-stream categorization, and positions MDWD to fill that gap. To show the dataset is usable, the authors train YOLO11, YOLO12, YOLO26, and RF-DETR models on a fixed 80/10/10 split; the best model, RF-DETR-M, reaches 94.49% mAP50 and 93.56% F1, and every model exceeds 83% F1. The practical point: municipal authorities and researchers get a reproducible benchmark for vision-based waste monitoring that reflects operational collection streams.

What carries the argument

The load-bearing object is the dataset itself, with its five-category taxonomy tied directly to Malta's color-coded collection system. The annotation protocol — manual bounding boxes, a '~20% visible' rule, single lead-annotator QA pass, and exclusion of burst captures and video frames — determines the label quality. The evaluation protocol fixes an 80/10/10 split, offline augmentation (ten variants per training image), and a shared training recipe for YOLO-family models, with RF-DETR trained on the same augmented data, so that reported scores measure dataset learnability rather than fine-tuned architecture differences.

What would settle it

Re-annotate a random subset of, say, 200 MDWD images with at least two independent annotators and compute inter-annotator agreement (e.g., IoU and category kappa); also run a duplicate/near-duplicate search (perceptual hash or embedding similarity) across the train/test split. If agreement is low or duplicates appear across splits, the benchmark scores are not a trustworthy signal.

Watch

Extended reading notes

Core claim

MDWD is a public benchmark of 3,697 street-level images annotated with 11,461 axis-aligned bounding boxes across five waste categories: Mixed Waste (black bags), Organic Waste (white bags), Recyclable Material (gray/green bags), Orange CMD (distinctive orange bags), and Other Waste (a long-tail residual class). The paper's central claim is that this dataset occupies a position none of the reviewed waste datasets does — street-level context, instance-level localization, and multiple operationally defined domestic waste streams — and that the benchmark results (RF-DETR-M: mAP50 94.49%, F1 93.56%; all models above 83% F1 on test) show the annotations are learnable across CNN-based, attention-ce

Load-bearing premise

The reported mAP and F1 numbers depend on two unverified premises: that the 11,461 manual labels are correct and consistent, and that the 80/10/10 split contains no duplicate or near-duplicate images of the same waste pile.

Editorial extensions

If this is right

  • A public, green-field benchmark now exists for instance-level municipal waste detection from street imagery, letting future work compare detectors on a common protocol.
  • Because RF-DETR and YOLO variants all exceed 83% test F1, the dataset is learnable enough that accuracy differences below a few points likely reflect architecture choice rather than annotation noise.
  • The weakest class, Other Waste (mAP50 93.2, recall 83.5), gives a concrete target for long-tail waste categories.
  • Validation-test gaps are small, so the split appears stable enough for model selection.
  • Pre-trained checkpoints (COCO for YOLO, Objects365 for RF-DETR) transfer well, meaning the task aligns with generic object detection priors.

Reading between the lines

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

  • If the dataset is adopted as a benchmark, the next test is whether detectors trained on it transfer to other cities with different bag colors and collection schemes; the paper does not address geographic transfer.
  • The absence of an inter-annotator agreement statistic means the 11,461 boxes' reliability is untested; a natural follow-up is a label-quality study on a random subset.
  • The fixed augmentation pipeline and single seed make the baseline reproducible, but also mean sensitivity to augmentation choices is unexplored.
  • The orange CMD class's high precision/low recall pattern suggests a deployment system could use it for targeted collection alerts rather than general detection.
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

2 major / 6 minor

Summary. This paper introduces MDWD, a street-level benchmark for municipal solid waste detection comprising 3,697 high-resolution images and 11,461 manually annotated instances across five classes that mirror Malta's color-coded municipal collection streams (Mixed Waste, Organic Waste, Recyclable Material, Orange CMD, and Other Waste). The authors report an 80/10/10 train–validation–test split, offline augmentation, and a cross-architecture baseline study covering YOLO11, YOLO12, YOLO26, and RF-DETR at several scales. RF-DETR-M achieves the strongest test performance (mAP50 94.49%, F1 93.56%), and all evaluated models exceed 83% test F1. The paper claims that no existing dataset simultaneously provides street-level imagery, instance-level bounding boxes, and categorization into multiple operationally defined domestic waste streams, and that the results demonstrate the dataset supports effective detector training across CNN-based, attention-centric, and transformer-based paradigms.

Significance. If the underlying annotations and split are reliable, MDWD is a genuinely useful resource: it addresses a real gap, is publicly released on Roboflow, uses an operationally grounded taxonomy, and provides a fixed split with consistent validation-to-test gaps across ten models. The cross-architecture comparison is a practical strength, and the small val-test deltas in Table IV are encouraging evidence of split stability. However, the benchmark's value rests on the correctness of 11,461 manual boxes and on the independence of the train/test partitions, neither of which is currently measured. The stress-test concern about annotation consistency lands: the absence of inter-annotator agreement or label-noise statistics means the numbers in Table IV cannot yet be read as a trustworthy benchmark signal. With added annotation-consistency and duplicate-image analyses, this would be a solid benchmark contribution.

major comments (2)
  1. [Section III.A / Table IV] The ground-truth boxes are the benchmark signal, but their correctness is asserted rather than measured. The QA pass by a single lead annotator is described, yet no inter-annotator agreement statistic or label-noise estimate is reported. The protocol's own decision rules are most difficult exactly where the paper acknowledges visual ambiguity: 'approximately 20% visible' is subjective, and the boundaries between Mixed Waste (black bags) and Recyclable Material (gray/green bags), as well as the residual Other Waste class, are noted to be visually similar or heterogeneous. If the lead annotator imposed idiosyncratic color or boundary preferences, every metric in Table IV and the Section IV.A claim that MDWD 'supports effective detector training' would be compromised. Add an inter-annotator agreement study on a stratified sample (e.g., per-class kappa) or a label-error audit; without it the
  2. [Section III.A / III.B] Split independence is asserted but not verified. The text states that each image depicts a distinct waste pile and that burst captures and video frames were excluded, but no duplicate or near-duplicate analysis is reported. Since images were collected opportunistically on foot and from moving vehicles, the same pile could plausibly appear in more than one image with a slightly different viewpoint; if any such near-duplicates cross the 80/10/10 boundary, the small validation-to-test gaps in Table IV would reflect leakage rather than generalization. Report a pairwise image-similarity screen (e.g., perceptual hashing or feature matching) across partitions, or provide acquisition metadata demonstrating no same-pile overlap. This is load-bearing for the benchmark's generalization claim.
minor comments (6)
  1. [Section III.C / Table IV] The sentence 'Detector architecture and model scale constitute the sole independent variables' is overstated because RF-DETR was trained on the Roboflow cloud platform defaults with Objects365 pretraining, while the YOLO models use custom Ultralytics settings with COCO pretraining. This conflates architecture with infrastructure, optimizer, and pretraining. The paper already acknowledges this in the following paragraph; please qualify the statement explicitly or soften it.
  2. [Table IV] All results appear to come from a single training run (YOLO seed 42; RF-DETR seed not reported). For a benchmark paper, reporting variance over two or three seeds, or at least noting the single-run status in the table caption, would let readers gauge the stability of the reported differences.
  3. [Section IV.B] The class-level analysis for RF-DETR-M is reported only as unnumbered prose (e.g., 'mAP50 98.3', 'false negative count 46'). A table with per-class mAP50, precision, recall, and counts would improve reproducibility and make the long-tail discussion easier to verify.
  4. [Section IV.A] Precision, recall, and F1 are computed with 'a fixed-threshold matching procedure (confidence and IoU thresholds of 0.50)', while mAP50 and mAP50:95 follow each framework's native COCO-style evaluation. Please clarify whether the fixed threshold is applied to the same detections used for mAP and state the confidence threshold explicitly; this affects comparability of the P/R/F1 columns.
  5. [Section III.B / Table II] Vertical flipping is applied with 50% probability. For street-level waste imagery this creates unnatural gravity orientation and may introduce artifacts not present in deployment. This is not blocking, but a brief justification or ablation would strengthen confidence in the augmentation choice.
  6. [Various] Minor textual issues: Table IV has a formatting error in the YOLO26-L Precision T column ('97.18' with no space before the next value), and the class distribution caption in Figure 1 reports percentages that should be cross-checked against the 11,461 total (e.g., 27.7% for Mixed Waste).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark numbers are measured on a held-out test split, and the dataset's categories and gap claim rest on external operational and literature anchors rather than on the paper's own outputs.

full rationale

The paper's derivation chain is self-contained as a benchmark contribution: it constructs a dataset with an externally anchored five-class taxonomy (Malta's color-coded municipal collection streams), performs manual annotation, fixes an 80/10/10 split before training, trains multiple detectors under a shared protocol, and reports measured test-set metrics. No fitted parameter is renamed as a prediction: the mAP/F1 values in Table IV are evaluations on a held-out test partition, not quantities solved for during the derivation. The central claim that MDWD 'supports effective detector training across a broad range of architectural paradigms' is directly supported by the reported test F1-scores, all above 83%, and is not a restatement of the dataset's definition. The gap claim is argued against named external datasets (TrashNet, TACO, GIGO, SODA, GVP, etc.), and while two citations involve overlapping authors (SODA [7] and the UAV litter review [8]), they are used as prior work and motivation, not as a load-bearing proof of the present results. The absence of an inter-annotator agreement statistic and the lack of duplicate-image analysis are reproducibility and validity concerns, but they do not constitute circularity: even if the labels were noisy, the reported results would still be measurements rather than conclusions forced by definition. There is no self-definitional step, no fitted input called a prediction, and no uniqueness theorem imported from the authors' prior work. Accordingly, the paper merits a circularity score of 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on (a) the correctness and consistency of the manual annotations, and (b) the independence of the train/test partitions; neither is demonstrated with statistics (no inter-annotator agreement, no duplicate analysis). The evaluation is a measurement, not a derivation, so the only free parameters are hand-chosen training and labeling rules. No invented entities in the physics sense; the five-class taxonomy is the closest construct, and it is anchored to an external operational scheme.

free parameters (4)
  • Motion blur augmentation strength = up to 40 px
    Section III.B: 'selected empirically based on observed validation performance improvement' - a hand-set value tuned on the validation split.
  • Annotation visibility threshold = 20% visible
    Section III.A: objects annotated only if approximately 20% of extent is visible; a hand-set rule that determines ground-truth instance counts.
  • Photometric augmentation ranges = hue ±19, sat ±29%, brightness ±17%, exposure ±10%
    Section III.B, Table II: fixed hand-set magnitudes that shape the training distribution.
  • Precision/Recall/F1 matching thresholds = confidence 0.50, IoU 0.50
    Section IV.A: fixed-threshold matching used for P/R/F1; affects the reported numbers.
assumptions (4)
  • domain assumption Manual annotations are correct ground truth
    Section III.A: labels from trained annotators consolidated by a single lead reviewer; no inter-annotator agreement is reported, so the 11,461 boxes are assumed correct and consistent.
  • domain assumption Train/test split is leakage-free (each image a distinct waste pile)
    Section III.A: burst captures and video frames excluded, 'each image depicting a distinct waste pile'; no duplicate or near-duplicate analysis is reported.
  • domain assumption COCO/Objects365 pretrained weights transfer to street-level waste imagery
    Section III.C: all models start from COCO or Objects365 checkpoints; transfer is assumed and never compared against training from scratch.
  • standard math COCO-style mAP and fixed-threshold P/R/F1 correctly summarize detector quality here
    Section IV: metrics follow framework-native COCO-style evaluation and fixed conf/IoU 0.50 matching; standard practice assumed without task-specific validation.
invented entities (1)
  • Five-class waste taxonomy with residual 'Other Waste' bucket independent evidence
    purpose: Operationally grounded annotation scheme matching Malta's color-coded streams; 'Other Waste' is a residual catch-all for bulky refuse and bottles that anchors the long tail.
    Categories refer to real, externally observable objects (colored bags, bottles, bulky refuse), so the scheme is not purely internal; however, 'Other Waste' is defined by exclusion rather than by a positive operational criterion, making its boundary partly ad hoc.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MDWD: A Street-Level Dataset for Municipal Solid Waste Detection in Dense Urban Environments." pith.science (2026). https://pith.science/paper/VZAY37FY

@misc{pith2026260800257,
  author       = {Pith},
  title        = {Pith review of: MDWD: A Street-Level Dataset for Municipal Solid Waste Detection in Dense Urban Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZAY37FY}},
  note         = {Machine review of arXiv:2608.00257}
}
read the original abstract

Automated visual monitoring of urban environments is a growing Computer Vision research area, but municipal solid waste detection remains under-represented in dedicated benchmark resources. Existing waste-related datasets predominantly address individual litter detection, aerial imagery, or image-level classification, and none simultaneously provide street-level imagery, instance-level localization, and categorization of domestic waste streams within a structured municipal collection context. This paper introduces the Maltese Domestic Waste Dataset (MDWD), a street-level benchmark comprising 3,697 high-resolution images and 11,461 manually annotated instances across five domestic waste categories representative of Malta's municipal collection system. The dataset captures substantial variation in location, illumination, object scale, occlusion, and urban context. To establish reproducible baselines, a cross-architecture benchmark is conducted across multiple generations of the YOLO family and a transformer-based detector. On the test set, RF-DETR-M achieves the strongest overall performance with an mAP50 of 94.49% and an F1-score of 93.56%, whilst smaller-capacity variants maintain competitive accuracy at substantially reduced parameter counts. These results indicate that MDWD supports effective training across both compact real-time detectors and transformer-based models, establishing a benchmark for future research in vision-based municipal waste monitoring.

Figures

Figures reproduced from arXiv: 2608.00257 by the authors.

Figure 1
Figure 1. Class distribution of the 11,461 annotated instances in the MDWD. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Representative MDWD training sample after offline mosaic aug [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 1 canonical work pages

  1. [1]

    Exploring big data applications in sustainable urban infrastructure: A review,

    D. V . Ogunkan and S. K. Ogunkan, “Exploring big data applications in sustainable urban infrastructure: A review,”Urban Governance, vol. 5, no. 1, pp. 54–68, 2025. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S2664328625000038

    2025

  2. [2]

    Augmenting urban planning with computer vision: A review of the state-of-the-art,

    R. Marasinghe, T. Yigitcanlar, X. Fu, and S. J. Quan, “Augmenting urban planning with computer vision: A review of the state-of-the-art,” Sustainable Cities and Society, vol. 138, p. 107209, 2026

    2026

  3. [3]

    Municipal waste statistics,

    Eurostat, “Municipal waste statistics,” Eurostat Statistics Explained, 2026, accessed: 2026-06-10. [Online]. Available: https://ec.europa.eu/e urostat/statistics-explained/index.php?title=Municipal_waste_statistics

    2026

  4. [4]

    Municipal solid waste management and its impact: a review,

    M. Adnan, A. Jha, S. Kumaret al., “Municipal solid waste management and its impact: a review,”International Journal of Advanced Research in Engineering and Technology (IJARET), vol. 11, no. 5, pp. 685–693, 2020

    2020

  5. [5]

    Performance analysis of yolo, faster r-cnn, and detr for automated personal protective equipment detection,

    R. Naufaldihanif, D. Kurniawan, and K. D. Tania, “Performance analysis of yolo, faster r-cnn, and detr for automated personal protective equipment detection,”Journal of Applied Informatics and Computing, vol. 9, no. 6, p. 3810–3820, Dec. 2025. [Online]. Available: https://jurnal.polibatam.ac.id/index.php/JAIC/article/view/11593

    2025

  6. [6]

    Garbage vulnerable point monitoring using iot and computer vision,

    R. Kumar, A. Lall, S. Chaudhari, M. Kale, and A. Vattem, “Garbage vulnerable point monitoring using iot and computer vision,” in2025 IEEE 11th World Forum on Internet of Things (WF-IoT), 2025, pp. 1–6

    2025

  7. [7]

    Soda: A dataset for small object detection in uav captured imagery,

    D. Pisani, D. Seychell, C. J. Debono, and M. Schembri, “Soda: A dataset for small object detection in uav captured imagery,” in2024 IEEE International Conference on Image Processing (ICIP), 2024, pp. 151–157

    2024

  8. [8]

    Litter detection from aerial imagery: A review of uav-based approaches and deep learning techniques,

    M. Bartolo, G. Hili, D. Seychell, M. Montebello, C. J. Debono, S. Formosa, and K. Makantasis, “Litter detection from aerial imagery: A review of uav-based approaches and deep learning techniques,” Multimedia Tools and Applications, vol. 85, p. 232, 2026. [Online]. Available: https://doi.org/10.1007/s11042-026-21440-1

    work page doi:10.1007/s11042-026-21440-1 2026

Show all 27 references
  1. [9]

    A dataset for detection and segmentation of underwater marine debris in shallow waters,

    A. Ðuraš, B. J. Wolf, A. Ilioudi, I. Palunko, and B. De Schutter, “A dataset for detection and segmentation of underwater marine debris in shallow waters,”Scientific Data, vol. 11, p. 921, 2024

    2024

  2. [10]

    Taco: Trash annotations in context for litter detection,

    P. F. Proença and P. Simões, “Taco: Trash annotations in context for litter detection,” 2020. [Online]. Available: https://arxiv.org/abs/2003.06975

    2020 arXiv

  3. [11]

    A multi-level approach to waste object segmentation,

    T. Wang, Y . Cai, L. Liang, and D. Ye, “A multi-level approach to waste object segmentation,”Sensors, vol. 20, no. 14, 2020. [Online]. Available: https://www.mdpi.com/1424-8220/20/14/3816

    2020

  4. [12]

    Classification of trash for recyclability status,

    M. Yang and G. Thung, “Classification of trash for recyclability status,” CS229 project report, vol. 2016, no. 1, p. 3, 2016

    2016

  5. [13]

    Bottle detection in the wild using low-altitude unmanned aerial vehicles,

    J. Wang, W. Guo, T. Pan, H. Yu, L. Duan, and W. Yang, “Bottle detection in the wild using low-altitude unmanned aerial vehicles,” in2018 21st International Conference on Information Fusion (FUSION), 2018, pp. 439–444

    2018

  6. [14]

    Gigo, garbage in, garbage out: An urban garbage classification dataset,

    M. Sukel, S. Rudinac, and M. Worring, “Gigo, garbage in, garbage out: An urban garbage classification dataset,” inMultiMedia Modeling, D.-T. Dang-Nguyen, C. Gurrin, M. Larson, A. F. Smeaton, S. Rudinac, M.- S. Dao, C. Trattner, and P. Chen, Eds. Cham: Springer International Pu...

    2023

  7. [15]

    Intelligent computer vision applications for automated urban infrastructure monitoring,

    M. Ashraf and R. Pavel, “Intelligent computer vision applications for automated urban infrastructure monitoring,”ISCSITR-International Journal of Computer Applications (ISCSITR-IJCA), vol. 7, no. 1, pp. 1–7, 2026. [Online]. Available: https://iscsitr.in/index.php/ISCSITR-IJC A...

    2026

  8. [16]

    Towards general urban monitoring with vision-language models: A review, evaluation, and a research agenda,

    A. Torneiro, J. D. Monteiro, P. Novais, P. Rangel Henriques, and N. F. Rodrigues, “Towards general urban monitoring with vision-language models: A review, evaluation, and a research agenda,”arXiv preprint arXiv:2510.12400, 2025. [Online]. Available: https://arxiv.org/abs/2510.12400

    2025

  9. [17]

    A multimodal vision: Language framework for intelligent detection and semantic interpretation of urban waste,

    V . M. Jonuzi and I. Mishkovski, “A multimodal vision: Language framework for intelligent detection and semantic interpretation of urban waste,”Informatics, vol. 13, no. 4, 2026. [Online]. Available: https://www.mdpi.com/2227-9709/13/4/57

    2026

  10. [18]

    Harnessing deep learning for sustainable waste classification: An innovative approach using mobilenetv2 on a diverse bag dataset,

    E. Jain and R. Kumar, “Harnessing deep learning for sustainable waste classification: An innovative approach using mobilenetv2 on a diverse bag dataset,” in2024 International Conference on Cybernation and Computation (CYBERCOM), 2024, pp. 335–339

    2024

  11. [19]

    Enhancing waste management with automated garbage bag classification with resnet50v2,

    ——, “Enhancing waste management with automated garbage bag classification with resnet50v2,” in2024 Eighth International Conference on Parallel, Distributed and Grid Computing (PDGC), 2024, pp. 150– 154

    2024

  12. [20]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,” 2024. [Online]. Available: https: //arxiv.org/abs/2410.17725

    2024 arXiv

  13. [21]

    Yolov12: Attention-centric real-time object detectors,

    Y . Tian, Q. Ye, and D. DOERMANN, “Yolov12: Attention-centric real-time object detectors,” inAdvances in Neural Information Processing Systems, D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, Eds., vol. 38. Curran Associates, Inc., 2025, pp. 78...

    2025

  14. [22]

    Ultralytics yolo26: Unified real-time end-to-end vision models,

    G. Jocher, J. Qiu, M. Liu, S. Lyu, F. C. Akyon, and M. E. Kalfaoglu, “Ultralytics yolo26: Unified real-time end-to-end vision models,” 2026. [Online]. Available: https://arxiv.org/abs/2606.03748

    2026 arXiv

  15. [23]

    Rf- detr: Neural architecture search for real-time detection transformers,

    I. Robinson, P. Robicheaux, M. Popov, D. Ramanan, and N. Peri, “Rf- detr: Neural architecture search for real-time detection transformers,”

  16. [24]

    Sam 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, J. Lei, T. Ma, B. Guo, A. Kalla, M. Marks, J. Greer, M. Wang, P. Sun, R. Rädle, T. Afouras, E. Mavroudi, K. Xu, T.-H. Wu, Y . Zhou, L. Momeni, R. Hazra, S. Ding, S. Va...

    2026 arXiv

  17. [25]

    V-jepa 2: Self-supervised video models enable understanding, prediction and planning,

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, Mojtaba, Komeili, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus, S. Arnaud, A. Gejji, A. Martin, F. R. Hogan, D. Dugas, P. Bojanowski, V . Khalidov, P. Labatut, F. Massa, M. Szafraniec, K. Krishnakumar, Y . Li, X. Ma,...

    2025 arXiv

  18. [26]

    Ranking regions of visual saliency in rgb-d content,

    D. Seychell and C. J. Debono, “Ranking regions of visual saliency in rgb-d content,” in2018 International Conference on 3D Immersion (IC3D), 2018, pp. 1–8

    2018

  19. [2026]

    Available: https://arxiv.org/abs/2511.09554

    [Online]. Available: https://arxiv.org/abs/2511.09554

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.