Pith. sign in

REVIEW 4 major objections 5 minor 51 references

Efficient Object Detection of Marine Debris using Pruned YOLO Model

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Channel pruning a fine-tuned YOLOv4 produces a marine-debris detector that runs at 19.4 FPS with 96.4% mAP, losing only 1.2 points against the full model.

desk verdict Plausible pruning tradeoff, but mAP numbers are not trustworthy without a disclosed evaluation protocol and a video-aware split; routine extension of prior work with no code/data. read the letter →

arxiv 2501.16571 v1 pith:54KXMKJ5 submitted 2025-01-27 cs.CV cs.AI

Abi Aryaza , Novanto Yudistira , Tibyani This is my paper · ORCID
classification cs.CVcs.AI
keywords marinedebrisobjectdetectionYOLOv4channelpruningautonomousunderwatervehiclesTrash-ICRA19networkslimming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that channel pruning a fine-tuned YOLOv4, rather than switching to a lighter network, is the practical way to get real-time marine-debris detection on hardware an autonomous underwater vehicle can carry. The authors compare six training schemes on the Trash-ICRA 19 dataset and report that pruning 50% of the channels of a pretrained and fine-tuned YOLOv4 raises speed from 15.19 to 19.4 frames per second while mean average precision (mAP) drops from 97.6% to 96.4%. Pruning only 20% reaches 97.55% mAP, slightly above the unpruned mosaic-augmented base of 97.45%. If these numbers hold, a low-GPU robot can keep nearly full detection accuracy while gaining real-time speed.

What carries the argument

The central mechanism is channel pruning via sparse regularization, a network-slimming technique: during training, L1 regularization is imposed on the scaling factors stored in batch-normalization layers, driving unimportant channels' factors toward zero so those channels can be removed after training. The paper applies this to the YOLOv4 one-stage detector, which uses a CSPDarknet53 backbone, a PANet neck, and a three-scale detection head, after fine-tuning on Trash-ICRA 19. The pruning ratio is the control knob; at 50% it cuts parameters from 63.95M to 17.66M and raises FPS from 15.19 to 19.4 while losing only 1.2 points of mAP.

What would settle it

Rerun the 50%-pruned model on the public Trash-ICRA 19 test split using a fixed protocol that names the overlap threshold, confidence threshold, NMS setting, batch size, and FPS measurement method; if the measured mAP is not close to 96.4%, or the pruned model is not faster than the unpruned baseline, the paper's central claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that the best efficiency gains come from pruning channels out of an already fine-tuned YOLOv4 detector, not from freezing layers or from switching to the lighter YOLOv4-tiny. The paper's headline result is that a 50%-pruned YOLOv4 reaches 96.4% mAP at 19.4 FPS, against 97.6% mAP at 15.19 FPS for the unpruned fine-tuned model, with the parameter count falling from 63.95M to 17.66M. At 20% pruning the mAP is 97.55%, a hair above the base. The same experiments show that freezing the backbone and neck together hurts accuracy sharply and that mosaic augmentation does not consistently help, so full fine-tuning of the pretrained network is the stronger starting point. The authors interpret these results as making YOLOv4 usable for real-time marine-debris detection even on a low-end mobile GPU.

Load-bearing premise

The load-bearing premise is that the reported mAP and FPS values are externally valid and comparable to prior work, even though the paper does not state the overlap threshold, confidence threshold, non-maximum-suppression settings, or FPS measurement conditions.

Editorial extensions

If this is right

  • A 50%-pruned YOLOv4 on Trash-ICRA 19 runs at 19.4 FPS, which is fast enough for real-time underwater garbage collection on low-GPU AUV hardware.
  • A 20%-pruned YOLOv4 achieves 97.55% mAP, slightly above the 97.45% mAP of the base model it was pruned from, so low pruning ratios can speed up the network without giving up accuracy.
  • Freezing the backbone and neck together drops the best mAP to below 88%, so full fine-tuning of a pretrained YOLOv4 is the stronger recipe before pruning.
  • YOLOv4-tiny is the fastest option among the tested schemes at up to 31.65 FPS but reaches only 94.4% mAP, leaving the pruned YOLOv4 as the middle-ground detector.

Reading between the lines

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

  • Inference: A fixed-protocol reproduction on Trash-ICRA 19 that states the overlap threshold, confidence threshold, NMS settings, and FPS measurement conditions would turn the reported efficiency gain into a directly transferable number for AUV designers.
  • Inference: If the pruning speed-up transfers to other underwater datasets, pruned YOLOv4 could become the default detector for low-power marine robots, since it keeps most of the full model's accuracy at a fraction of the parameter count.
  • Inference: The FPS curve in the pruning table has a non-monotonic point at 70% pruning, so probing ratios between 50% and 60% is a natural search for an even better operating point than the selected 50%.
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 / 5 minor

Summary. The paper proposes and evaluates several YOLOv4 variants for marine debris detection on the Trash-ICRA 19 dataset, comparing training from scratch, pretraining, layer freezing, YOLOv4-tiny, and channel pruning. Its central claim is that a 50%-pruned pretrained YOLOv4 achieves 96.4% mAP at 19.4 FPS, increasing the base model's FPS from 15.19 with only a roughly 1.2% mAP drop, making the model suitable for real-time detection on low-GPU AUVs.

Significance. If the reported tradeoff is reproducible and evaluated under a protocol comparable to prior work, the result is a practically useful data point for deploying one-stage detectors on low-resource underwater robots. The paper also provides a systematic comparison of freezing and augmentation strategies on a publicly available marine debris dataset. However, the evaluation protocol is under-specified, no code or data are released, and the reported mAP values are far above published results on the same dataset; these issues must be resolved before the central claim can be accepted. The paper does not present machine-checked proofs or parameter-free derivations; its strength would lie in reproducible empirical evidence, which is currently missing.

major comments (4)
  1. [4.6 and Conclusion] The headline claim of a 1.2% mAP drop is computed against the wrong baseline. Scheme 6 prunes the pretrained model with mosaic augmentation, whose base mAP is 97.45% per Table 2, but the conclusion compares 96.4% to the 97.6% without-mosaic baseline. Using the correct base, the drop is about 1.0 percentage point (97.45% to 96.46%). The abstract and conclusion should be corrected, and because the reported numbers come from single runs without error bars, the difference between 97.45% and 96.46% should be interpreted with appropriate caution.
  2. [4, Table 10] The evaluation protocol is not specified. The paper never states the IoU threshold, confidence threshold, NMS settings, or the exact evaluation code used to compute mAP. This is load-bearing because Table 10 compares the proposed mAP of 96-97% with prior work on the same dataset (e.g., Fulton et al. 47.9%, Majchrowska et al. 7.3%) without any evidence that the same protocol was applied. The authors should provide their evaluation script or report results under a standard protocol (e.g., COCO-style or VOC-style mAP) and, ideally, run a same-protocol comparison on the same test split.
  3. [1 and 4] The train/test split appears to be frame-level, which risks temporal leakage. The paper states in Section 1 that Trash-ICRA 19 images were captured every 3 frames per second, producing many similar images, yet the split is described only by counts (5719/1144/820) with no mention of video-aware grouping, temporal gaps, or deduplication. If test frames are near-duplicates of training frames, the high mAP could reflect memorization rather than generalization. The authors should either perform a video-aware split or measure and report frame-level similarity between train and test sets.
  4. [3.3 and 4.6] The channel-pruning pipeline is under-specified. Section 3.3 describes network slimming at a high level, but Section 4.6 gives no details on the sparse-training epochs, the L1 regularization coefficient, whether the pruning ratio is global or per-layer, how channels are selected, or the post-pruning fine-tuning schedule. Without these details and the code, the speed/accuracy tradeoff in Table 6 is not reproducible. The authors should provide the full pruning recipe and, ideally, release the code and trained models.
minor comments (5)
  1. [Figure 17] The captions report "mAP: 99.99%" for three example images; mAP is a dataset-level metric and cannot be computed on three images. These captions should be removed or replaced with per-image confidence scores.
  2. [References] Reference [20] is incomplete; the article title is missing from the citation.
  3. [Throughout] The text uses "Schema" and "Scheme" interchangeably (e.g., "Schema 1" in Section 4.1 versus "Scheme 6" elsewhere); please standardize the terminology.
  4. [Table 6] The unusual FPS drop at 70% pruning is explained by an excessive number of predicted bounding boxes, but this explanation is not quantified. Reporting the average number of detections per image at each pruning ratio would make the explanation verifiable.
  5. [4.7] The selection of 20% and 50% as highlighted pruning ratios was made after scanning 10-90% in Table 6; this post-hoc selection should be acknowledged, and the authors should note the lack of multiple runs for each ratio.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the pruned-model results are direct experimental measurements, and the only self-citations are motivational, not load-bearing.

full rationale

The paper does not derive results from fitted equations, self-defined quantities, or an author-imported uniqueness theorem. It trains YOLOv4 variants and reports measured mAP, FPS, and parameter counts in Tables 1-10; the central claim that channel pruning raises FPS while modestly lowering mAP is an empirical comparison of those measured values. The channel-pruning method follows the standard network-slimming approach of Liu et al. rather than a self-defined construction, so the outcome is not equivalent to the input by definition. The 20% and 50% pruning ratios are selected after scanning 10-90% in Table 6, but the reported values are measurements of trained models, not predictions forced by the selection rule. References [17] and [18] are authored by the same group, but they only motivate the mosaic-augmentation and medical-imaging discussions; the paper's own experiments independently evaluate both mosaic augmentation and pruning, so these citations are not load-bearing. Concerns about the undisclosed IoU/NMS/FPS protocol, the frame-level train/test split given 3 fps capture, and the mismatch between pruning the with-mosaic model and comparing to the without-mosaic base are validity and reproducibility issues, not circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

This is an empirical ML paper with no mathematical derivation. The free parameters are experimental hyperparameters that directly affect the reported trade-off. The axioms are standard domain assumptions for transfer learning and dataset usage, but the evaluation-protocol assumption is load-bearing given the unexplained mAP gap with prior work.

free parameters (2)
  • Pruning ratio (50% for the headline model) = 50%
    Selected post hoc after scanning 10-90% pruning ratios; the paper highlights 20% and 50% as superior without a pre-specified criterion.
  • Sparse regularization coefficient for channel pruning = Not reported
    Network slimming requires choosing the L1 penalty strength on BN scaling factors; this value affects which channels are pruned and is not stated.
assumptions (3)
  • domain assumption Trash-ICRA 19 annotations and split are accurate and representative for marine debris detection.
    The paper relies on the dataset's three classes and its fixed split without questioning label quality or whether the split matches prior benchmark protocols.
  • domain assumption Pretrained COCO weights transfer effectively to underwater debris images.
    Fine-tuning pretrained YOLOv4 is assumed to yield the reported high mAP; no analysis of domain shift is provided.
  • domain assumption The implementation of channel pruning follows the standard network-slimming procedure and is correctly applied to YOLOv4.
    The paper describes the general idea (L1 regularization on BN scales) but does not specify the pruning tool, fine-tuning schedule, or handling of shortcut connections, which are known to be nontrivial for YOLO architectures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Object Detection of Marine Debris using Pruned YOLO Model." pith.science (2026). https://pith.science/paper/54KXMKJ5

@misc{pith2026250116571,
  author       = {Pith},
  title        = {Pith review of: Efficient Object Detection of Marine Debris using Pruned YOLO Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/54KXMKJ5}},
  note         = {Machine review of arXiv:2501.16571}
}
read the original abstract

Marine debris poses significant harm to marine life due to substances like microplastics, polychlorinated biphenyls, and pesticides, which damage habitats and poison organisms. Human-based solutions, such as diving, are increasingly ineffective in addressing this issue. Autonomous underwater vehicles (AUVs) are being developed for efficient sea garbage collection, with the choice of object detection architecture being critical. This research employs the YOLOv4 model for real-time detection of marine debris using the Trash-ICRA 19 dataset, consisting of 7683 images at 480x320 pixels. Various modifications-pretrained models, training from scratch, mosaic augmentation, layer freezing, YOLOv4-tiny, and channel pruning-are compared to enhance architecture efficiency. Channel pruning significantly improves detection speed, increasing the base YOLOv4 frame rate from 15.19 FPS to 19.4 FPS, with only a 1.2% drop in mean Average Precision, from 97.6% to 96.4%.

Figures

Figures reproduced from arXiv: 2501.16571 by the authors.

Figure 1
Figure 1. Example of image dataset Given the two types of object detection, opting for a one-stage detector is more suitable for Autonomous Underwater Vehicles (AUVs). AUVs are designed for the purpose of collecting trash in the ocean, necessitating real-time detection capabilities. YOLO[20] [21] and Single-Shot Detector (SSD)[22] [23] are several of notable object detection models that can be employed for real-time object de… view at source ↗
Figure 2
Figure 2. Object Detection Evolution[33] YOLOv4 represents an advancement of YOLO, offering the flexibility to choose its backbone. By selecting CSPDarknet53 as the backbone, YOLOv4 achieves superior mAP and FPS compared to CSPResNeXt-50 and EfficientNet- B0[35]. CSPDarknet53 is an enhanced iteration of Darknet-53, the original backbone for YOLOv3. This enhancement involves the incorporation of CSPNet (Cross Stage Partial Net… view at source ↗
Figure 3
Figure 3. YOLOv4 Architecture 3.2. Loss Function In the architecture of YOLOv4, the selection of a loss function plays a pivotal role in optimizing the model’s performance. YOLOv4 incorporates four types of loss functions to efficiently train the network. These encompass the CIoU loss [41], object confidence loss, no-object confidence loss, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Channel Pruning Diagram Channel pruning removing channels that are considered less crucial. The importance of each channel can be determined through sparse regularization, which identifies channels that are infrequently utilized. This is achieved by inserting L1-regula…
Figure 5
Figure 5. Figure 5: Schema 1 of Training YOLOv4 from Scratch [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Loss value on Schema 1 The loss values for both training sets decrease rapidly until epoch 200, after which the reduction in loss occurs more gradually. The model converges after epoch 400. The sudden drop in loss at epoch 201 is attributed to the adjustment of the lea…
Figure 7
Figure 7. Figure 7: Schema 2 of fine-tuning pretrained YOLOv4 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Loss Value of YOLOv4 Pretrained The [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Schema 3 of fine-tuning YOLOv4 with frozen backbone layer [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Loss Value of YOLOv4 Freeze Backbone Layer [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Schema 4 of YOLOv4 with frozen backbone+neck layer [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Loss Value of YOLOv4 Freeze Backbone and Neck Layer [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Schema 5 of training YOLOv4-tiny from scratch [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: Loss Value of YOLOv4-tiny [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Schema 6 of fine-tuning pruned-YOLOv4 [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 16
Figure 16. Figure 16: YOLOv4-Pruned mAP Result Graph [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 17
Figure 17. Figure 17: Prediction Result 30 [PITH_FULL_IMAGE:figures/full_fig_p030_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 31 canonical work pages

  1. [1]

    M. Z. Elamin, K. N. Ilmi, T. Tahrirah, Y. A. Zarnuzi, Y. C. Suci, D. R. Rahmawati, D. M. D. P., R. Kusumaardhani, R. A. Rohmawati, P. A. Bhagaskara, I. F. Nafisa, Analysis of waste management in the village of disanah, district of sreseh sampang, madura, JURNAL KESEHATAN LINGKUNGAN 10 (2018) 368. doi:10.20473/jkl.v10i4.2018.368-375. URL https://e-journal....

    work page doi:10.20473/jkl.v10i4.2018.368-375 2018

  2. [2]

    K. L. H. dan Kehutanan, Capaian kinerja pengelolaan sampah (2023)

    work page 2023

  3. [3]

    Oceanic, A

    N. Oceanic, A. Administration, What is marine debris? (4 2021)

    work page 2021

  4. [4]

    Cherson, The marine debris research, prevention and reduction act: A policy analysis (2005)

    A. Cherson, The marine debris research, prevention and reduction act: A policy analysis (2005). doi:10.13140/RG.2.2.26619.54562. URL https://www.researchgate.net/publication/344306086

    work page arXiv 2005

  5. [5]

    T. K. N. P. S. Laut, Sampah laut (2023)

    work page 2023

  6. [6]

    Hyakudome, Design of autonomous underwater vehicle (2011)

    T. Hyakudome, Design of autonomous underwater vehicle (2011). URL www.intechopen.com

    work page 2011

  7. [7]

    Y. C. Chang, S. K. Hsu, C. H. Tsai, Sidescan sonar image processing: Correcting brightness variation and patching gaps, Journal of Marine Science and Technology 18 (2010) 785–789. doi:10.51400/2709- 6998.1935

    work page arXiv 2010

  8. [8]

    Wiley, T

    V. Wiley, T. Lucas, Computer vision and image processing: A paper review, International Journal of Artificial Intelligence Research 2 (2018) 22. doi:10.29099/ijair.v2i1.42

    work page doi:10.29099/ijair.v2i1.42 2018

Show all 51 references
  1. [9]

    Cosido, A

    O. Cosido, A. Iglesias, A. Galvez, R. Catuogno, M. Campi, L. Teran, E. Sainz, Hybridization of convergent photogrammetry, computer vision, and artificial intelligence for digital documentation of cultural heritage-a case study: The magdalena palace, Institute of Electrical and...

    2014 doi

  2. [10]

    Z. Zou, K. Chen, Z. Shi, Y. Guo, J. Ye, Object detection in 20 years: A survey (5 2019), Proceedings of the IEEE 111 (3) (2023) 257–276. doi:10.1109/JPROC.2023.3238524

    2023

  3. [11]

    L. Du, R. Zhang, X. Wang, Overview of two-stage object detection algorithms, Vol. 1544, Institute of Physics Publishing, 2020. doi:10.1088/1742-6596/1544/1/012033

    2020 doi

  4. [12]

    Carranza-Garc´ıa, J

    M. Carranza-Garc´ıa, J. Torres-Mateo, P. Lara-Ben´ıtez, J. Garc´ıa- Guti´errez, 33 On the performance of one-stage and two-stage object detectors in autonomous vehicles using camera data, Remote Sensing 13 (2021) 1–23. doi:10.3390/rs13010089

    2021 doi

  5. [13]

    Janakiramaiah, G

    B. Janakiramaiah, G. Kalyani, A. Karuna, L. V. N. Prasad, M. Krishna, P. V. Potluri, G. K. V. R. Siddhartha, Military object detection in defence using multi-level capsule networks (3 2021)

    2021

  6. [14]

    Aralikatti, J

    A. Aralikatti, J. Appalla, S. Kushal, G. S. Naveen, S. Lokesh, B. S. Jayasri, Real- time object detection and face recognition system to assist the visually impaired, Vol. 1706, IOP Publishing Ltd, 2020. doi:10.1088/1742- 6596/1706/1/012149

    2020 doi

  7. [15]

    Balasubramaniam, S

    A. Balasubramaniam, S. Pasricha, Object detection in autonomous vehicles: Status and open challenges (2022)

    2022

  8. [16]

    Y. Hu, G. Liu, Z. Chen, J. Guo, Object detection algorithm for wheeled mobile robot based on an improved yolov4, Applied Sciences 12 (9) (2022). doi:10.3390/app12094769. URL https://www.mdpi.com/2076-3417/12/9/4769

    2022 doi

  9. [17]

    A. S. Nugraha, Y. Novanto, B. Rahayudi, Supervised virtual-to-real domain adaptation for object detection task using yolo , in: 2024 IEEE Conference on Artificial Intelligence (CAI), IEEE Computer Society, Los Alamitos, CA, USA, 2024, pp. 1359–1364. doi:10.1109/CAI59869.2024.0...

    2024

  10. [18]

    Tibyani, Penerapan region growing pada analisis citra digital untuk pendeteksian sel-sel kanker rahim, Tesis S2 Teknik Elektro (2005)

    2005

  11. [19]

    Srivastava, A

    S. Srivastava, A. V. Divekar, C. Anilkumar, I. Naik, V. Kulkarni, V. Pattabiraman, Comparative analysis of deep learning image detection algorithms, Journal of Big Data 8 (12 2021). doi:10.1186/s40537-021- 00434- w

    2021 doi

  12. [20]

    H. Yang, P. Liu, Y. Hu, J. Fu, , Microsystem Technologies 27 (2021) 1837–1844. doi:10.1007/s00542-019-04694-8. URL https://link.springer.com/10.1007/s00542-019-04694-8

    2021 doi

  13. [21]

    Y. H. Liao, J. G. Juang, Automatic marine debris inspection, Aerospace 10 (1 2023). doi:10.3390/aerospace10010084

    2023 doi

  14. [22]

    Tata, S.-J

    G. Tata, S.-J. Royer, O. Poirion, J. Lowe, A robotic approach towards quantifying epipelagic bound plastic using deep visual models (5 2021). URL http://arxiv.org/abs/2105.01882 34

    2021 arXiv

  15. [23]

    B. Xue, B. Huang, W. Wei, G. Chen, H. Li, N. Zhao, H. Zhang, An efficient deep-sea debris detection method using deep neural networks, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14 (2021) 12348–12360. doi:10.1109/JSTARS.2021.3130238

    2021

  16. [24]

    Algorry, A

    A. Algorry, A. Giles Garcia, G. Wolfmann, Real-time object detection and classification of small and similar figures in image processing, 2017, pp. 516–519. doi:10.1109/CSCI.2017.87

    2017 doi

  17. [25]

    Benjdira, T

    B. Benjdira, T. Khursheed, A. Koubaa, A. Ammar, K. Ouni, Car detection using unmanned aerial vehicles: Comparison between faster r-cnn and yolov3, in: 2019 1st International Conference on Unmanned Vehicle Systems-Oman (UVS), 2019, pp. 1–6. doi:10.1109/UVS.2019.8658300

    2019

  18. [26]

    M. S. A. B. Rosli, I. S. Isa, M. I. F. Maruzuki, S. N. Sulaiman, I. Ahmad, Underwater animal detection using yolov4, Institute of Electrical and Electronics Engineers Inc., 2021, pp. 158–163. doi:10.1109/ICCSCE52189.2021.9530877

    2021

  19. [27]

    M. Tian, X. Li, S. Kong, L. Wu, J. Yu, A modified yolov4 detection method for a vision-based underwater garbage cleaning robot, Frontiers of Information Technology & Electronic Engineering 23 (2022) 1217– 1228. doi:10.1631/FITEE.2100473. URL https://link.springer.com/10.1631/F...

    2022 doi

  20. [28]

    Fulton, J

    M. Fulton, J. Hong, M. J. Islam, J. Sattar, Robotic detection of marine litter using deep visual detection models, IEEE, 2019, pp. 5752–5758. doi:10.1109/ICRA.2019.8793975. URL https://ieeexplore.ieee.org/document/8793975/

    2019

  21. [29]

    Majchrowska, A

    S. Majchrowska, A. Miko-lajczyk, M. Ferlin, Z. Klawikowska, M. A. Plantykow, A. Kwasigroch, K. Majek, Waste detection in pomerania: non- profit project for detecting waste in environments, Waste Management 138 (2022) 274–284. doi:https://doi.org/10.1016/j.wasman.2021.12.001 UR...

    2022 doi

  22. [30]

    S´anchez-Ferrer, J

    A. S´anchez-Ferrer, J. J. Valero-Mas, A. J. Gallego, J. Calvo- Zaragoza, An experimental study on marine debris location and recognition using object detection, Pattern Recognition Letters (4 2023). doi:10.1016/j.patrec.2022.12.019

    2023 doi

  23. [31]

    M. Tian, X. Li, S. Kong, L. Wu, J. Yu, Pruning-based yolov4 algorithm for underwater gabage detection, Vol. 2021-July, IEEE Computer Society, 2021, pp. 4008–4013. doi:10.23919/CCC52363.2021.9550592. 35

    2021

  24. [32]

    Tajbakhsh, J

    N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B. Gotway, J. Liang, Convolutional neural networks for medical image analysis: Full training or fine tuning?, IEEE Transactions on Medical Imaging 35 (2016) 1299–1312. doi:10.1109/TMI.2016.2535302

    2016

  25. [33]

    Diwan, G

    T. Diwan, G. Anirudh, J. V. Tembhurne, Object detection using yolo: challenges, architectural successors, datasets and applications, Multimedia Tools and Applications 82 (2023) 9243–9275. doi:10.1007/s11042- 022- 13644-y

    2023 doi

  26. [34]

    Hosang, R

    J. Hosang, R. Benenson, B. Schiele, Learning non-maximum suppression , in: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE Computer Society, Los Alamitos, CA, USA, 2017, pp. 6469–6477. doi:10.1109/CVPR.2017.685. URL https://doi.ieeecomputersociety....

    2017 doi

  27. [35]

    Mahasin, I

    M. Mahasin, I. A. Dewi, Comparison of cspdarknet53, cspresnext-50, and efficientnet-b0 backbones on yolo v4 as object detector, International Journal of Engineering, Science & InformationTechnology 2 (2022) 64–72. doi:10.52088/ijesty.v1i4.291. URL https://doi.org/10.52088/ijes...

    2022 doi

  28. [36]

    Bochkovskiy, C.-Y

    A. Bochkovskiy, C.-Y. Wang, H.-Y. M. Liao, Yolov4: Optimal speed and accuracy of object detection (4 2020). URL http://arxiv.org/abs/2004.10934

    2020 arXiv

  29. [37]

    Misra, Mish: A self regularized non-monotonic activation function, in: British Machine Vision Conference, 2020

    D. Misra, Mish: A self regularized non-monotonic activation function, in: British Machine Vision Conference, 2020. URL https://api.semanticscholar.org/CorpusID:221113156

    2020

  30. [38]

    Qiang, W

    Z. Qiang, W. Yuanyu, Z. Liang, Z. Jin, L. Yu, L. Dandan, Research on real-time reasoning based on jetson tx2 heterogeneous acceleration yolov4, Institute of Electrical and Electronics Engineers Inc., 2021, pp. 455–459. doi:10.1109/ICCCBDA51879.2021.9442515

    2021

  31. [39]

    S. Liu, L. Qi, H. Qin, J. Shi, J. Jia, Path aggregation network for instance segmentation, 2018, pp. 8759–8768 doi:10.1109/CVPR.2018.00913

    2018

  32. [40]

    Redmon, A

    J. Redmon, A. Farhadi, Yolov3: An incremental improvement (04 2018). arXiv:1804.02767. 36

    2018 arXiv

  33. [41]

    Zheng, P

    Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, D. Ren, Distance-iou loss: Faster and better learning for bounding box regression, Vol. 34, 2020, pp. 12993– 13000. doi:10.1609/aaai.v34i07.6999

    2020 doi

  34. [42]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks (2012). URL http://code.google.com/p/cuda-convnet/

    2012

  35. [43]

    Simonyan, A

    K. Simonyan, A. Zisserman, Very deep convolutional networks for large- scale image recognition (2015). arXiv:1409.1556

    2015 arXiv

  36. [44]

    Szegedy, W

    C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, A. Rabinovich, Going deeper with convolutions (2014). arXiv:1409.4842

    2014 arXiv

  37. [45]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation (2015). arXiv:1505.04597

    2015 arXiv

  38. [46]

    K. He, G. Gkioxari, P. Doll´ar, R. Girshick, Mask r-cnn, in: 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980– 2988. doi:10.1109/ICCV.2017.322

    2017 doi

  39. [47]

    M. M. Pasandi, M. Hajabdollahi, N. Karimi, S. Samavi, Modeling of pruning techniques for simplifying deep neural networks, in: 2020 International Conference on Machine Vision and Image Processing (MVIP), 2020, pp. 1–

    2020

  40. [48]

    doi:10.1109/MVIP49855.2020.9116891

    2020

  41. [49]

    Z. Liu, J. Li, Z. Shen, G. Huang, S. Yan, C. Zhang, Learning efficient convolutional networks through network slimming (8 2017). URL http://arxiv.org/abs/1708.06519

    2017 arXiv

  42. [50]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition (2015). arXiv:1512.03385. URL https://arxiv.org/abs/1512.03385

    2015 arXiv

  43. [51]

    T. Lin, M. Maire, S. J. Belongie, L. D. Bourdev, R. B. Girshick, J. Hays, P. Perona, D. Ramanan, P. Doll’a r, C. L. Zitnick, Microsoft COCO: common objects in context, CoRR abs/1405.0312 (2014). arXiv:1405.0312. URL http://arxiv.org/abs/1405.0312

    2014 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.