Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Policy-Driven Transfer Learning in Resource-Limited Animal Monitoring

T0 review · 4 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 An upper-confidence-bound bandit selects the best pre-trained animal detector from 16 candidates while running far fewer model executions than exhaustive evaluation.

desk verdict A practical model-selection idea undermined by an internal numerical inconsistency: same selected model, different reported F1. read the letter →

arxiv 2509.10995 v1 pith:PVTY5HSX submitted 2025-09-13 cs.CV

Nisha Pillai , Aditi Virupakshaiah , Harrison W. Smith , Amanda J. Ashworth , Prasanna Gowda , Phillip R. Owens , Adam R. Rivers , Bindu Nanduri
show 1 more author
Mahalingam Ramkumar
This is my paper
classification cs.CV
keywords animaldetectiontransferlearningreinforcementupperconfidenceboundmodelselectionobjectUAVimagerylimiteddata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a reinforcement-learning agent guided by the upper confidence bound (UCB) algorithm can solve a practical bottleneck: choosing which pre-trained object detector to use for animal monitoring when labeled data and compute are scarce. Instead of running all 16 COCO-pretrained models on every UAV image, the agent selects one model per image, updates a reward estimate, and converges on the best candidate. On sheep and cattle imagery, the UCB selector reaches F1 0.718, above the 0.690 of brute-force evaluation and 0.517 of a consensus baseline, while using a fraction of the model executions. The authors conclude that adaptive bandit selection can replace exhaustive search in resource-limited transfer learning.

What carries the argument

The upper confidence bound (UCB) bandit algorithm, Algorithm 2, with exploration constant C=0.1, together with the reward-penalty update in Algorithm 3: for each image, the selected model gains Qa += b for matched boxes, Qa -= g−b for false negatives, and Qa -= d−b for false positives. These accumulated rewards guide both which model to try next and which model is returned as best via arg max_a Qa. The machinery's job is to balance trying untested models against exploiting models that have scored well, so the search stops well short of exhaustive evaluation.

What would settle it

Compute the Spearman rank correlation between each model's cumulative reward Q and its F1 score on the training images; if the model with the highest Q is not among the top-F1 models, the proxy misranks. A direct falsifier: find a candidate dataset where the UCB selector's arg-max reward model has lower F1 than another candidate that scored lower reward, on held-out data.

Watch

Extended reading notes

Core claim

The central claim is that a UCB-driven reward-penalty scheme, which selects one model per image based on cumulative rewards and an exploration bonus, identifies the same high-performing model (RTDETRx) as exhaustive search, but with fewer executions and with a higher test F1. The authors frame this as a transfer-learning pipeline where pre-trained COCO weights are used without fine-tuning on the target animal data. This shows that a simple bandit can do the model-selection work that normally requires testing every candidate on every image.

Load-bearing premise

The reward update in Algorithm 3 (adding b for correct detections, subtracting false negatives and false positives) is assumed to rank candidate models in the same order as the harmonic-mean F1 score, but the paper provides no evidence that this linear proxy tracks F1.

Editorial extensions

If this is right

  • The UCB selector runs one model per image instead of all 16, so the computational cost of model selection drops by roughly an order of magnitude, with the reported test F1 improving from 0.690 to 0.718.
  • The consensus baseline, which fuses predictions from multiple models, performs worse (0.517) than picking a single good model, suggesting that fusion is not a useful fallback on this data.
  • RTDETRx, a real-time transformer detector, is the top model in all approaches, indicating that transformer-based detectors can excel on limited RGB animal data.
  • The approach is intended for researchers with constrained compute and small datasets, who can use the selector to pick a pre-trained model without exhaustive testing.
  • Because the reward is model-global, the final selected model can be deployed directly or fine-tuned further.

Reading between the lines

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

  • If the reward proxy's ranking of models does not match F1 ordering on a given dataset, the selector could return a model with high recall but low precision; a rank-correlation check between cumulative reward and F1 would be a cheap validation before trusting the selection.
  • The compute savings should grow as the candidate pool grows, since UCB prunes low-reward models after a few negative trials; on pools larger than 16, the gap versus brute-force should widen.
  • The current formulation selects one global model; it could be extended to per-image selection, where different models handle different scenes, but that would change the reward semantics and needs separate validation.
  • The reported numbers rest on a small test set (~14 images after a 90:10 split), so the F1 gap between methods should be read with caution until repeated runs or confidence intervals are reported.
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 / 6 minor

Summary. The paper addresses pre-trained object-detection model selection for UAV-based cattle/sheep monitoring with limited labeled data. It evaluates 16 COCO-pretrained models on 137 RGB cattle images and proposes a reinforcement-learning/UCB selection algorithm that runs one model per image, using a reward based on true-positive, false-negative, and false-positive counts. The method is compared with brute-force exhaustive RL and a consensus fusion baseline. The authors claim that the UCB approach achieves F1=0.718, higher than brute-force RL (0.690) and consensus (0.517), while executing far fewer model evaluations. The paper also identifies RTDETRx as the best model in both the UCB and brute-force runs.

Significance. If substantiated, the framework would be a practical contribution to model selection in resource-constrained animal monitoring, where exhaustive evaluation of many detectors is often prohibitive. The paper has strengths: it provides explicit pseudocode for the algorithm, uses external ground-truth annotations (no circularity), and documents the data-collection and annotation pipeline. However, the central empirical claim is not currently supported. There is an unexplained inconsistency in Table II (same selected model, different F1), no statistical significance evidence, and no validation that the linear reward proxy ranks models by F1. These issues are load-bearing because the abstract, Section VI, and the conclusion all rest on the numerical superiority and efficiency of the UCB method.

major comments (4)
  1. [Section VI, Table II] The brute-force RL and UCB rows both report RTDETRx as the selected model, yet their test F1 values differ (0.690 vs 0.718). With fixed pre-trained weights and fixed thresholds (ρ≥0.5, τ≥0.5), the same model evaluated on the same test set must yield identical precision, recall, and F1. This discrepancy implies that the two methods used different test splits, different thresholds, or different image subsets. The paper does not state whether the 90:10 partition is identical across methods or how the four repetitions were assigned. This invalidates the headline comparison until the evaluation protocol is unified and the numbers are recomputed.
  2. [Section V, Algorithm 3] The reward update is Q_a += b; Q_a -= g-b; Q_a -= d-b, which simplifies to Q_a ← TP - FN - FP. The final model is selected by argmax Q, not by F1, and no evidence is provided that this linear proxy orders models the same way as the harmonic-mean F1. Since the brute-force baseline also uses this reward, the F1 values in Table II may correspond to a reward-maximizing model rather than an F1-maximizing model. Please show the correlation between reward ranking and F1 ranking, or replace the selection criterion with one that directly optimizes F1.
  3. [Section VI] No variance or significance measures are reported. The text says each evaluation was run four times and averaged, but no standard deviations or confidence intervals are given. With only 137 images, a 90:10 split leaves roughly 14 test images, so an F1 difference of 0.028 is likely within sampling noise. Report per-run results, error bars, and a paired significance test using the same splits for all methods.
  4. [Sections IV.C and VI] The paper is internally inconsistent about which model exhaustive evaluation selects. Section IV.C and Figure 3 identify YOLO11x as the highest-performing model under ρ≥0.5, τ≥0.5; Section VI says the top classifiers are 'YOLOv11x and RT-DETR'; Table II says RTDETRx is the top model across all approaches. Because UCB and brute-force both select RTDETRx, but the preliminary exhaustive evaluation reportedly picked YOLO11x, the reader cannot tell which model is actually best, so the claim that UCB matches the exhaustive selection is unsupported. Clarify and reconcile these statements.
minor comments (6)
  1. [Algorithm 2] Typo: 'explorationCostant' should be 'explorationConstant'.
  2. [Table I and Section IV.C] The table row labeled 'τ≥0.01' conflicts with the text saying the IoU threshold ρ was reduced to 0.01. Correct the label and the corresponding description.
  3. [Abstract / Section III.C] The acronym 'UA V' should be 'UAV'.
  4. [Section IV.A] The class label set includes 'person' twice and contains the typo 'giraf f e'. Clean up the set and notation.
  5. [Throughout] Model names are inconsistent: RT-DETR vs RTDETRx and YOLOv11x vs YOLO11x. Use a single naming convention.
  6. [Abstract and Section VI] The claim of 'significantly less computational time' is qualitative; no execution counts or wall-clock times are reported. Provide quantitative efficiency measures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: model selection rewards are computed from external ground-truth labels, and the reported F1 is an independent evaluation metric.

full rationale

The paper's derivation chain is not circular. The RL/UCB framework (Algorithm 3) computes rewards from counts of correct detections (b), false negatives (g-b), and false positives (d-b), all derived from comparing model predictions against manually annotated ground-truth bounding boxes. The final reported metric is F1-score, computed separately from the same external ground truth. No parameter is fitted to the target F1 result: the exploration constant C=0.1 and thresholds ρ=0.5, τ=0.5 are fixed prior to evaluation. The paper contains no load-bearing self-citations and does not invoke a uniqueness theorem or an ansatz from the authors' prior work. The reward function is an ad hoc linear proxy and there is an unexplained inconsistency in Table II (the same selected model, RTDETRx, is reported with different F1 values under brute-force RL and UCB), but these are correctness/reproducibility concerns, not circularity. The prediction of the best model is not equivalent to the reported F1 by construction. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on four hand-chosen parameters (C, rho, tau, reward weights) and on domain assumptions about the suitability of COCO pre-trained models and annotation accuracy. There are no invented physical or algorithmic entities.

free parameters (4)
  • Exploration constant C = 0.1
    Hand-chosen in Algorithm 3 to control the exploration-exploitation trade-off in UCB.
  • IoU threshold rho = 0.5
    Set in Algorithm 3; defines when a predicted box matches a ground truth box.
  • Confidence threshold tau = 0.5
    Set in Algorithm 3; filters out low-confidence detections before reward computation.
  • Reward weights = +1 for TP, -1 for FN, -1 for FP
    Ad hoc linear reward in Algorithm 3; not derived from F1 or any stated objective.
assumptions (3)
  • domain assumption Pre-trained COCO models can detect animals in aerial drone images without fine-tuning
    The entire evaluation relies on it; Section IV applies models with pre-trained weights directly to cattle images at 30m altitude.
  • domain assumption Manual ground truth annotations are complete and accurate
    Section III-D describes manual annotation as ground truth without inter-annotator agreement or quality checks.
  • standard math Standard UCB convergence properties hold for selecting the best model
    The paper assumes UCB will identify the optimal model over time (Section V), relying on standard bandit theory without proof in this setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Policy-Driven Transfer Learning in Resource-Limited Animal Monitoring." pith.science (2026). https://pith.science/paper/PVTY5HSX

@misc{pith2026250910995,
  author       = {Pith},
  title        = {Pith review of: Policy-Driven Transfer Learning in Resource-Limited Animal Monitoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PVTY5HSX}},
  note         = {Machine review of arXiv:2509.10995}
}
read the original abstract

Animal health monitoring and population management are critical aspects of wildlife conservation and livestock management that increasingly rely on automated detection and tracking systems. While Unmanned Aerial Vehicle (UAV) based systems combined with computer vision offer promising solutions for non-invasive animal monitoring across challenging terrains, limited availability of labeled training data remains an obstacle in developing effective deep learning (DL) models for these applications. Transfer learning has emerged as a potential solution, allowing models trained on large datasets to be adapted for resource-limited scenarios such as those with limited data. However, the vast landscape of pre-trained neural network architectures makes it challenging to select optimal models, particularly for researchers new to the field. In this paper, we propose a reinforcement learning (RL)-based transfer learning framework that employs an upper confidence bound (UCB) algorithm to automatically select the most suitable pre-trained model for animal detection tasks. Our approach systematically evaluates and ranks candidate models based on their performance, streamlining the model selection process. Experimental results demonstrate that our framework achieves a higher detection rate while requiring significantly less computational time compared to traditional methods.

Figures

Figures reproduced from arXiv: 2509.10995 by the authors.

Figure 1
Figure 1. Overview of the transfer learning approach. We create [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Examples of object detection results on RGB images. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Using pre-trained weights from the COCO dataset, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Overview of the Reinforcement learning (RL) Process [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 4 linked inside Pith

  1. [1]

    Automati- cally identifying, counting, and describing wild animals in camera-trap images with deep learning,

    M. S. Norouzzadeh, A. Nguyen, M. Kosmala, A. Swan- son, M. S. Palmer, C. Packer, and J. Clune, “Automati- cally identifying, counting, and describing wild animals in camera-trap images with deep learning,”Proceedings of the National Academy of Sciences, 2018

    2018

  2. [2]

    Complexity in ecology and con- servation: mathematical, statistical, and computational challenges,

    J. L. Green, A. Hastings, P. Arzberger, F. J. Ayala, K. L. Cottingham, K. Cuddington, F. Davis, J. A. Dunne, M.-J. Fortin, L. Gerberet al., “Complexity in ecology and con- servation: mathematical, statistical, and computational challenges,”BioScience, 2005

    2005

  3. [3]

    K ¨uhl,Best practice guidelines for the surveys and monitoring of great ape populations

    H. K ¨uhl,Best practice guidelines for the surveys and monitoring of great ape populations. IUCN, 2008

    2008

  4. [4]

    Exploring artificial intelligence for applications of drones in forest ecology and manage- ment,

    A. Buchelt, A. Adrowitzer, P. Kieseberg, C. Gollob, A. Nothdurft, S. Eresheim, S. Tschiatschek, K. Stampfer, and A. Holzinger, “Exploring artificial intelligence for applications of drones in forest ecology and manage- ment,”Forest Ecology and Management, 2024

    2024

  5. [5]

    A comprehensive review of deep learning approaches for animal detection on video data

    P. Kumar, S. Luo, and K. Shaukat, “A comprehensive review of deep learning approaches for animal detection on video data.”International Journal of Advanced Com- puter Science & Applications, vol. 14, no. 11, 2023

    2023

  6. [6]

    Using computer vision, image analysis and uavs for the auto- matic recognition and counting of common cranes (grus grus),

    A. Chen, M. Jacob, G. Shoshani, and M. Charter, “Using computer vision, image analysis and uavs for the auto- matic recognition and counting of common cranes (grus grus),”Journal of Environmental Management, 2023

    2023

  7. [7]

    Review on methods used for wildlife species and individual identification,

    T. Petso, R. S. Jamisola Jr, and D. Mpoeleng, “Review on methods used for wildlife species and individual identification,”European Journal of Wildlife Research, vol. 68, no. 1, p. 3, 2022

    2022

  8. [8]

    Towards the fully automated monitoring of ecological communities,

    M. Besson, J. Alison, K. Bjerge, T. E. Gorochowski, T. T. Høye, T. Jucker, H. M. Mann, and C. F. Clements, “Towards the fully automated monitoring of ecological communities,”Ecology Letters, 2022

    2022

Show all 40 references
  1. [9]

    Synchronizing object detection: Applications, advancements and existing challenges,

    M. T. Hosain, A. Zaman, M. R. Abir, S. Akter, S. Mur- salin, and S. S. Khan, “Synchronizing object detection: Applications, advancements and existing challenges,” IEEE Access, 2024

    2024

  2. [10]

    Deep learning in multi-object detection and tracking: state of the art,

    S. K. Pal, A. Pramanik, J. Maiti, and P. Mitra, “Deep learning in multi-object detection and tracking: state of the art,”Applied Intelligence, 2021

    2021

  3. [11]

    A comprehensive review of convolutional neural networks for defect detection in industrial applications,

    R. Khanam, M. Hussain, R. Hill, and P. Allen, “A comprehensive review of convolutional neural networks for defect detection in industrial applications,”IEEE Access, 2024

    2024

  4. [12]

    A reinforcement learn- ing (rl)-based hybrid method for ground penetrating radar (gpr)-driven buried object detection,

    M. N. Alpdemir and M. Sezgin, “A reinforcement learn- ing (rl)-based hybrid method for ground penetrating radar (gpr)-driven buried object detection,”Neural Computing and Applications, vol. 36, no. 14, pp. 8199–8219, 2024

    2024

  5. [13]

    Deep transfer learning with fuzzy ensemble approach for the early detection of breast cancer,

    S. Sannasi Chakravarthy, N. Bharanidharan, V . Vinoth Kumar, T. Mahesh, M. S. Alqahtani, and S. Guluwadi, “Deep transfer learning with fuzzy ensemble approach for the early detection of breast cancer,”BMC Medical Imaging, 2024

    2024

  6. [14]

    A weighted ensemble transfer learning approach for melanoma classification from skin lesion images,

    H. Meswal, D. Kumar, A. Gupta, and S. Roy, “A weighted ensemble transfer learning approach for melanoma classification from skin lesion images,”Mul- timedia Tools and Applications, 2024

    2024

  7. [15]

    Hybrid models for classifying histological images: An associa- tion of deep features by transfer learning with ensemble classifier,

    C. I. De Oliveira, M. Z. do Nascimento, G. F. Roberto, T. A. Tosta, A. S. Martins, and L. A. Neves, “Hybrid models for classifying histological images: An associa- tion of deep features by transfer learning with ensemble classifier,”Multimedia Tools and Applications, 2024

    2024

  8. [16]

    Efficient automation of neural network design: A survey on dif- ferentiable neural architecture search,

    A. Heuillet, A. Nasser, H. Arioui, and H. Tabia, “Efficient automation of neural network design: A survey on dif- ferentiable neural architecture search,”ACM Computing Surveys, vol. 56, no. 11, pp. 1–36, 2024

    2024

  9. [17]

    Medical neural architecture search: Survey and taxonomy,

    H. Benmeziane, I. Hamzaoui, Z. Cherif, and K. El Maghraoui, “Medical neural architecture search: Survey and taxonomy,” inInternational Joint Conference on Artificial Intelligence, 2024

    2024

  10. [18]

    Laser: Learning to adaptively select reward models with multi-armed bandits,

    D. Nguyen, A. Prasad, E. Stengel-Eskin, and M. Bansal, “Laser: Learning to adaptively select reward models with multi-armed bandits,” 2024

    2024

  11. [19]

    Detrs beat yolos on real-time object detection,

    W. Lv, S. Xu, Y . Zhao, G. Wang, J. Wei, C. Cui, Y . Du, Q. Dang, and Y . Liu, “Detrs beat yolos on real-time object detection,” 2023

    2023

  12. [20]

    Processing coastal imagery with agisoft metashape professional edition, version 1.6—structure from motion workflow documen- tation,

    J.-S. R. Over, A. C. Ritchie, C. J. Kranenburg, J. A. Brown, D. D. Buscombe, T. Noble, C. R. Sherwood, J. A. Warrick, and P. A. Wernette, “Processing coastal imagery with agisoft metashape professional edition, version 1.6—structure from motion workflow documen- tation,” US Ge...

    2021

  13. [21]

    Computer Vision Annotation Tool (CV AT),

    CV AT.ai Corporation, “Computer Vision Annotation Tool (CV AT),” Nov. 2023

    2023

  14. [22]

    Microsoft COCO: common objects in context,

    T. Lin, M. Maire, S. J. Belongie, L. D. Bourdev, R. B. Girshick, J. Hays, P. Perona, D. Ramanan, P. Doll’a r, and C. L. Zitnick, “Microsoft COCO: common objects in context,”CoRR, vol. abs/1405.0312, 2014

    2014 arXiv

  15. [23]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” 2017

    2017

  16. [24]

    Jing Qiu, “

    A. C. ”Glenn Jocher, “Jing Qiu, “”ultralytics”,” 2023

    2023

  17. [25]

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

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779–788

    2016

  18. [26]

    Advancing crowd object detection: A review of yolo, cnn and vits hybrid approach,

    M. A. M. Ali, T. Aly, A. T. Raslan, M. Gheith, and E. A. Amin, “Advancing crowd object detection: A review of yolo, cnn and vits hybrid approach,”Journal of Intelligent Learning Systems and Applications, 2024

    2024

  19. [27]

    Yolov8: A novel object detection algorithm with enhanced performance and ro- bustness,

    R. Varghese and M. Sambath, “Yolov8: A novel object detection algorithm with enhanced performance and ro- bustness,” in2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS). IEEE, 2024, pp. 1–6

    2024

  20. [28]

    Yolov10: Real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding, “Yolov10: Real-time end-to-end object detection,”arXiv preprint arXiv:2405.14458, 2024

    2024 arXiv

  21. [29]

    Super-gradients,

    S. Aharon, Louis-Dupont, Ofri Masad, K. Yurkova, Lotem Fridman, Lkdci, E. Khvedchenya, R. Rubin, N. Bagrov, B. Tymchenko, T. Keren, A. Zhilko, and Eran- Deci, “Super-gradients,” 2021

    2021

  22. [30]

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

    T. Cheng, L. Song, Y . Ge, W. Liu, X. Wang, and Y . Shan, “Yolo-world: Real-time open-vocabulary object detection,”arXiv preprint arXiv:2401.17270, 2024

    2024 arXiv

  23. [31]

    Mask r-cnn,

    K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, “Mask r-cnn,” inProceedings of the IEEE international confer- ence on computer vision, 2017, pp. 2961–2969

    2017

  24. [32]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recog- nition, 2016, pp. 770–778

    2016

  25. [33]

    Liver segmentation in ct imaging with enhanced mask region-based convolutional neural networks,

    X. Chen, X. Wei, M. Tang, A. Liu, C. Lai, Y . Zhu, and W. He, “Liver segmentation in ct imaging with enhanced mask region-based convolutional neural networks,”An- nals of translational medicine, vol. 9, no. 24, 2021

    2021

  26. [34]

    Vehicle de- tection through instance segmentation using mask r-cnn for intelligent vehicle system,

    A. Ojha, S. P. Sahu, and D. K. Dewangan, “Vehicle de- tection through instance segmentation using mask r-cnn for intelligent vehicle system,” in2021 5th international conference on intelligent computing and control systems (ICICCS). IEEE, 2021, pp. 954–959

    2021

  27. [35]

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

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,”Advances in neural information processing systems, vol. 28, 2015

    2015

  28. [36]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications,

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” 2017

    2017

  29. [37]

    Fcos: Fully convolutional one-stage object detection

    Z. Tian, C. Shen, H. Chen, and T. He, “Fcos: Fully convolutional one-stage object detection.” 2019

    2019

  30. [38]

    Light-weight retinanet for object detection,

    Y . Li and F. Ren, “Light-weight retinanet for object detection,”arXiv preprint arXiv:1905.10011, 2019

    1905 arXiv

  31. [39]

    Ssd: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” inECCV 2016: 14th European Conference. Springer, 2016

    2016

  32. [40]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan, “Very deep convolutional networks for large-scale image recognition,” 2014

    2014

Pith tools

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