An Analysis of Layer-Freezing Strategies for Enhanced Transfer Learning in YOLO Architectures

Ana M. Bernardos; Andrzej D. Dobrzycki; Jos\'e R. Casar

REVIEW 4 major objections 6 minor 37 references

An Analysis of Layer-Freezing Strategies for Enhanced Transfer Learning in YOLO Architectures

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read There is no universal optimal freezing strategy for YOLO transfer learning; the data decides.

desk verdict Useful practical benchmark of YOLO freezing, but the headline rankings rest on single runs and one glaring internal inconsistency; treat the optimal-strategy recommendations as provisional. read the letter →

arxiv 2509.05490 v1 pith:CPQ3SVHT submitted 2025-09-05 cs.CV cs.AI

Andrzej D. Dobrzycki , Ana M. Bernardos , José R. Casar This is my paper

classification cs.CVcs.AI MSC 68T07

keywords YOLOlayerfreezingtransferlearningobjectdetectionUAVinfrastructuremonitoringmAP@50GPUmemoryoptimizationGrad-CAM

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 layer freezing in modern YOLO object detectors is not a one-size-fits-all choice: the best number of blocks to freeze depends on the target dataset's class balance, object scale, and augmentation. Running YOLOv8 and YOLOv10 variants on four UAV infrastructure-monitoring datasets, it finds that freezing either the first four feature-extraction blocks or the entire backbone usually reaches mAP@50 scores close to or better than full fine-tuning, while cutting GPU memory use substantially. Freezing most of the network fails badly on one heavily augmented single-class dataset, revealing a trainable-parameter capacity floor. The practical payoff is a rule: preserve general features for multi-class scenes, freeze less for extreme class imbalance, and do not freeze aggressively when augmentation shifts the input distribution.

What carries the argument

The operational object is the block-level decomposition of YOLOv8/YOLOv10 into three freezing regions: blocks 0-3 (pure feature extraction before the first concatenation), blocks 0-8 (the full backbone), and 22 or 23 blocks (most of the backbone plus neck/head fusion). Freezing these regions defines FR1, FR2, and FR3. The comparison is carried by four measurements taken across four model scales and four datasets: mAP@50 and mAP@50:95 for accuracy, peak GPU memory and training time for efficiency, L2 norms of gradients over training for convergence behavior, and Grad-CAM activation maps for what the frozen model attends to. The L2 gradient statistics distinguish chaotic high-variance optimiza

What would settle it

Take a YOLOv8n and a YOLOv10n model and run FR2 versus full fine-tuning on a heavily augmented, single-class dataset without extreme class imbalance and with at least 3,000 training images. If FR2 falls more than about 1.5% mAP@50 short of full fine-tuning while the paper's own datasets do not, the claim that data properties determine the best freeze depth is in trouble. A second check: re-run the Electric Substation comparison with FR1 redefined as freezing blocks 1-4 or 0-7; if the ranking changes materially, the architectural-boundary assumption is doing the work.

Watch

Extended reading notes

Core claim

On its own terms, the claim is that transfer learning in YOLOv8 and YOLOv10 has no universal freeze-depth optimum. Across four UAV infrastructure datasets, freezing blocks 0-3 (FR1) or the full backbone blocks 0-8 (FR2) yields mAP@50 scores within about 1.5% of the best configuration while cutting peak GPU memory by 25-68% in the recommended configurations, and in some cases beating full fine-tuning outright. Freezing 22/23 blocks (FR3) saves the most memory but fails catastrophically on the Bird's Nest dataset: YOLOv10n drops from 0.629 to 0.055 mAP@50, while the same FR3 remains workable on the other three datasets. Gradient L2 norms track these outcomes, with FR1/FR2 showing moderate stab

Load-bearing premise

The empirical rankings assume that the chosen freeze boundaries - after blocks 0-3, after the 9-block backbone, and after 22/23 blocks - are the functionally meaningful cuts in both architectures; if blocks 0-3 are not purely domain-agnostic or the 9-block boundary is not the real backbone, the comparison mischaracterizes what freezing preserves.

Editorial extensions

If this is right

Backbone freezing (FR2) is the recommended default for multi-class infrastructure scenes like InsPLAD-det and Common-VALID: it keeps mAP@50 within roughly 1.2-1.4% of the best model while cutting GPU memory by about 30-57%.
For extreme class imbalance (Electric Substation), freezing only the first four blocks (FR1) preserves adaptation capacity in mid-to-late layers and is the recommended trade-off.
Aggressive freezing (FR3) should be avoided on heavily augmented or distribution-shifted single-class data; on Bird's Nest it cuts mAP@50 by 57-91% depending on model variant.
Small models under aggressive freezing leave too few trainable parameters (below roughly 1-1.5 million) and fail on augmented data, so freeze depth must be chosen relative to model scale as well as data.
Monitoring gradient L2 norms during early training can flag poor freeze choices: an FR3-to-fine-tuning gradient ratio below 0.6 indicates likely convergence failure before training completes.

Reading between the lines

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

The paper stops short of a closed-form decision rule, but its results imply one: classify the target dataset by class imbalance, augmentation strength, and object scale, then map those attributes to freeze depth (FR1, FR2, or FR3). This mapping is testable on new datasets.
The 0.6 early-gradient-ratio threshold suggests a cheap diagnostic: run the first roughly 20 training steps under candidate freeze strategies, compare gradient ratios, and commit to the cheapest strategy whose ratio stays above threshold. The paper reports the signal but does not build the selector.
The Bird's Nest failure attributes the collapse to heavy augmentation plus undersized trainable capacity; a direct test would freeze FR3 on the same dataset without augmentation. If the collapse disappears, augmentation shift, not texture difficulty, is the causal driver.
Synthetic data (Common-VALID) behaved like real data under freezing, suggesting that backbone freezing could be paired with synthetic pretraining to stretch scarce real-world labels in infrastructure domains; the paper does not claim this pairing.

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.

Desk Editor's Note

read the letter

Quick take: this is a genuinely useful benchmark, not a conceptual breakthrough. It gives practitioners a wide, transparent map of how freezing depth interacts with model size and dataset type in YOLOv8/YOLOv10, and the qualitative message—no single optimal strategy, moderate freezing often matches full fine-tuning—is sensible and consistent with earlier YOLOv5 work. But the quantitative rankings in Table 8 rest on single-run measurements, and the paper contains an internal inconsistency big enough to undermine the specific 'optimal' recommendations.

What's new: a systematic sweep over two modern YOLO versions, four model sizes, and four datasets with different characteristics (class imbalance, object size, augmentation), plus gradient L2-norm curves, Grad-CAM, full result tables, and code/data on HuggingFace. That is real work with real reproducibility value. The GPU memory savings (up to ~28% or more) and the Bird's Nest failure under aggressive freezing are concrete, useful findings.

Where it gets soft: the stress-test note is correct. Table 6 lists v8s-9b on Common-VALID at 0.809 mAP@50; Appendix A Table A1 lists the same configuration at 0.881. Seven points is far outside the 1.5% tolerance used to select 'optimal' configurations. One of these numbers is wrong, and either way it means the tables weren't cross-checked. Second, every configuration was run exactly once (seed 42). Many strategy differences are 1-3 mAP points, which is within plausible stochastic variation, so claims like 'FR2 is optimal for InsPLAD' or 'FR1 for Electric Substation' are not statistically supported. The gradient-ratio threshold of 0.6 in Appendix C is estimated and validated on the same data, a mild circularity; it is fine as a descriptive observation, not as a 'reliable indicator'. The block-boundary assumption is reasonable but unvalidated; it is a secondary concern compared to the run-to-run variability.

Bottom line: the paper is honest about its limitations and will be useful to practitioners deciding where to start with YOLO freezing. It does not overclaim much in the text, but Table 8 does overclaim relative to the evidence. It deserves a serious referee—send it out—but the referee should require either a fix of the internal inconsistency plus some variability estimate (even 2-3 seeds), or a rewrite that clearly labels the recommendations as provisional single-run observations.

Referee Report

4 major / 6 minor

Summary. This manuscript presents an empirical study of layer-freezing strategies for YOLOv8 and YOLOv10 object detectors. For the nano, small, medium, and large variants, it compares training from scratch, full fine-tuning, freezing the first 4 blocks (FR1), freezing the 9-block backbone (FR2), and freezing 22/23 blocks (FR3) on four infrastructure-monitoring datasets (InsPLAD-det, Electric Substation, Common-VALID, Bird's Nest). It reports mAP@50, mAP@50:95, GPU memory usage, and training time, and it supplements the accuracy/efficiency tables with gradient L2-norm analyses and Grad-CAM visualizations. The central empirical claim is that there is no universally optimal freezing strategy: FR1 or FR2 often matches or beats full fine-tuning at lower GPU cost, while aggressive FR3 fails on the heavily augmented Bird's Nest dataset. The paper includes full result tables, appendices with additional freezing variants, and links to public code and data.

Significance. If the results are reproducible, the paper would provide useful, actionable guidance for practitioners deploying YOLO detectors under resource constraints. Its strengths are the breadth of the experimental grid, the reporting of complete tables, the release of code and data, and the combination of accuracy/efficiency metrics with gradient and Grad-CAM diagnostics. The conclusions, however, rest on single-run comparisons whose stability is not demonstrated, and at least one internal numerical inconsistency affects a headline result. The paper is therefore a potentially valuable empirical contribution, but the central ranking claim needs stronger statistical support before it can be accepted.

major comments (4)

[§5.1, Table 5; §6.1, Table 6; Appendix A, Table A1] Every configuration is run once with seed 42 (§5.1, Table 5), but the central dataset-dependent ranking is built from these single runs. There is an internal inconsistency: Table 6 reports v8s-9b on Common-VALID as mAP@50=0.809, while Appendix A, Table A1 reports the same model/dataset configuration as 0.881. This 0.072 difference is roughly five times the 1.5% tolerance used to select recommendations in Table 8. Please correct the discrepancy if it is a typo, re-check all tables, and provide multi-seed runs with means and error bars or confidence intervals. Without this, the rankings underlying the main claim are not statistically supported.
[§6.1, Table 8] The optimal-configuration selection uses a free 'performance tolerance threshold' of 1.5% and then chooses the minimal-GPU configuration among those within tolerance. No justification or sensitivity analysis is given for this threshold. Since many configurations in Tables 6–7 are separated by less than 1.5%, and each was run once, the recommendations ('FR2 optimal for InsPLAD/Common-VALID', 'FR1 optimal for Electric Substation', etc.) may be arbitrary. Provide error bars, bootstrap confidence intervals, or a sensitivity sweep over the tolerance value to demonstrate that the recommendations are robust.
[Appendix C] The claims that gradient-norm ratios below 0.6 'reliably indicate potential convergence failure' and that practitioners should freeze less than 50% of parameters are derived from exactly the two datasets used to illustrate them (Bird's Nest failure vs. InsPLAD success). This is post-hoc fitting rather than independent validation. Please label these as exploratory hypotheses or validate them on held-out configurations/datasets, and soften the 'reliably indicate' language.
[Abstract; §7; Tables 6 and 8] The reported GPU-savings figures are mutually inconsistent. The abstract states savings 'up to 28%', Section 7 states FR1 and FR2 reduce GPU memory 'by up to 28% and 44%', but values in Table 6 imply FR2 reductions up to ~57% (e.g., v8l-9b: 4724 vs. 10873 MB) and Table 8 reports a 68% saving for Bird's Nest. Since resource savings are a central contribution, these numbers must be reconciled and the reference baseline for each percentage stated explicitly.

minor comments (6)

[§6.1] There is a duplicated sentence: 'In contrast, approaches that only freeze the backbone layers...' appears twice verbatim in the same paragraph.
[Table 4] The table layout is difficult to read: the 'Approach' column values are not visually aligned with the variant rows, and the rounding conventions for the frozen-parameter percentages should be stated.
[Throughout] The typography 'YOLOvlO' is inconsistent and appears to be a rendering artifact; unify to 'YOLOv10' or match the notation used by the original YOLOv10 paper.
[§3.1; §5.1] The assumption that blocks 0–3 are 'pure feature extraction' before any concatenation and that 9-block backbone is the semantically meaningful unit for freezing is presented as architectural fact but is not empirically validated. Consider adding a small ablation or at least an explicit caveat that the FR1/FR2/FR3 comparisons depend on this boundary choice.
[§4.4; §5.1] Section 5.1 says data augmentation is intentionally disabled during training, yet the Bird's Nest training set is constructed with heavy augmentation. Clarify that the augmentation is applied offline in the dataset construction rather than online during training.
[Appendix A, Tables A1–A2] These 'constrained experiment' tables are useful, but it is not stated whether these runs use exactly the same seed and hyperparameters as the main experiments, which is relevant given the discrepancy noted in the major comments.

Circularity Check

1 steps flagged · score 2.0 of 10

No substantive circularity: central claims are measured outcomes; only minor post-hoc threshold fitting in Appendix C.

fitted input called prediction [Appendix C, final paragraph and Figure A5 caption]
"The Bird's Nest curves remain below the empirical 0.6 threshold, indicating restricted learning capacity. ... while monitoring gradient ratios during early training, as values below 0.6 reliably indicate potential convergence failure."

The 0.6 gradient-ratio threshold is extracted from the same two cases it is used to validate: InsPLAD-Det initial ratios are 0.68-0.70 (success) and Bird's Nest ratios stay below 0.6 (failure). After placing the cut between these observed curves, the paper restates 'values below 0.6' as a reliable indicator of convergence failure. This is in-sample separation presented as a predictive rule, not an independently tested prediction. It is a mild form of fitted-input-called-prediction, and it is confined to an appendix; the main dataset-dependent mAP conclusions do not rely on this threshold.

full rationale

The paper is an empirical study rather than a derivation: its central conclusions are direct readings of measured mAP@50, GPU usage, and training time across Tables 6-8. The claim that there is no universal optimal freezing strategy is a summary of observed rankings, not an equation that reduces to its inputs. The architectural decomposition into blocks 0-3, 0-8, and 22/23 is an assumption grounded in the YOLO literature (cited to [24]) and is stated before the experiments; it is not an output derived from the data. There is no load-bearing self-citation: references such as [24] are independent architectural reviews, and no 'uniqueness theorem' or prior author work is used to force the freezing choices. The only mildly circular element is the 0.6 gradient-ratio threshold in Appendix C, which is chosen after inspecting the two datasets it is then used to classify. That is a post-hoc heuristic, not part of the main accuracy/efficiency comparisons. The single-run design and the internal discrepancy between Table 6 (0.809) and Table A1 (0.881) for v8s-9b on Common-VALID are serious reproducibility concerns, but they are measurement-quality issues, not circularity. Overall circularity is low.

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

No new particles, forces, or abstract entities are introduced. The paper's contributions are empirical measurements and practical guidelines, not a new theoretical construct.

free parameters (2)

Performance tolerance threshold = 1.5%
Used in Table 8 to define the pool of acceptable configurations; chosen after seeing the results.
Gradient ratio critical threshold = 0.6
Used in Appendix C as an early predictor of training failure; derived from the same Bird's Nest and InsPLAD runs.

assumptions (2)

domain assumption Blocks 0-3 of YOLOv8/YOLOv10 form pure feature extraction layers and blocks 0-8 the full backbone, so freezing these exact boundaries preserves transferable visual features.
Section 3.1 and 3.2 assert this architectural decomposition, and the FR1/FR2/FR3 freezing boundaries depend on it.
domain assumption SGD hyperparameters (LR 1e-2, momentum 0.937, etc.) and the fixed training budget (1000 epochs with early stopping) provide a fair comparison across freezing strategies.
Section 5.1: if the chosen hyperparameters interact differently with frozen layers, the conclusions about strategy ranking could change.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Analysis of Layer-Freezing Strategies for Enhanced Transfer Learning in YOLO Architectures." pith.science (2026). https://pith.science/paper/CPQ3SVHT

@misc{pith2026250905490,
  author       = {Pith},
  title        = {Pith review of: An Analysis of Layer-Freezing Strategies for Enhanced Transfer Learning in YOLO Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPQ3SVHT}},
  note         = {Machine review of arXiv:2509.05490}
}

read the original abstract

The You Only Look Once (YOLO) architecture is crucial for real-time object detection. However, deploying it in resource-constrained environments such as unmanned aerial vehicles (UAVs) requires efficient transfer learning. Although layer freezing is a common technique, the specific impact of various freezing configurations on contemporary YOLOv8 and YOLOv10 architectures remains unexplored, particularly with regard to the interplay between freezing depth, dataset characteristics, and training dynamics. This research addresses this gap by presenting a detailed analysis of layer-freezing strategies. We systematically investigate multiple freezing configurations across YOLOv8 and YOLOv10 variants using four challenging datasets that represent critical infrastructure monitoring. Our methodology integrates a gradient behavior analysis (L2 norm) and visual explanations (Grad-CAM) to provide deeper insights into training dynamics under different freezing strategies. Our results reveal that there is no universal optimal freezing strategy but, rather, one that depends on the properties of the data. For example, freezing the backbone is effective for preserving general-purpose features, while a shallower freeze is better suited to handling extreme class imbalance. These configurations reduce graphics processing unit (GPU) memory consumption by up to 28% compared to full fine-tuning and, in some cases, achieve mean average precision (mAP@50) scores that surpass those of full fine-tuning. Gradient analysis corroborates these findings, showing distinct convergence patterns for moderately frozen models. Ultimately, this work provides empirical findings and practical guidelines for selecting freezing strategies. It offers a practical, evidence-based approach to balanced transfer learning for object detection in scenarios with limited resources.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 27 canonical work pages

[1]

Real-Time Flying Object Detection with YOLOv8

Reis, D.; Kupec, J.; Hong, J.; Daoudi, A. Real-Time Flying Object Detection with YOLOv8. arXiv 2024, arXiv:2305.09972. [CrossRef]

arXiv 2024
[2]

A YOLO-Based Traffic Counting System

Lin, J.P.; Sun, M.T. A YOLO-Based Traffic Counting System. In Proceedings of the 2018 Conference on Technologies and Applications of Artificial Intelligence (TAAI), Taichung, Taiwan, 30 November-2 December 2018; pp. 82-85

work page 2018
[3]

YOLO V3 + VGG16-based automatic operations monitoring and analysis in a manufacturing workshop under Industry 4.0

Yan, J.; Wang, Z. YOLO V3 + VGG16-based automatic operations monitoring and analysis in a manufacturing workshop under Industry 4.0. J. Manuf Syst. 2022, 63, 134-142. [CrossRef]

work page 2022
[4]

YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss

Maji, D.; Nagori, S.; Mathew, M.; Poddar, D. YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss. arXiv 2022, arXiv:2204.06806. [CrossRef]

work page Pith review arXiv 2022
[5]

A Comprehensive Survey on Transfer Learning

Zhuang, F.; Qi, Z.; Duan, K.; Xi, D.; Zhu, Y.; Zhu, H.; Xiong, H.; He, Q. A Comprehensive Survey on Transfer Learning. Proc. IEEE 2019, 109, 43-76. [CrossRef]

work page 2019
[6]

How transferable are features in deep neural networks? arXiv 2014, arXiv:1411.1792

Yosinski, J.; Clune, J.; Bengio, Y.; Lipson, H. How transferable are features in deep neural networks? arXiv 2014, arXiv:1411.1792. [CrossRef]

arXiv 2014
[7]

Microsoft COCO: Common Objects in Context

Lin, T.; Maire, M.; Belongie, S.J.; Bourdev, L.D.; Girshick, R.B.; Hays, J.; Perona, P.; Ramanan, D.; Dollar, P.; Zitnick, C.L. Microsoft COCO: Common Objects in Context. arXiv 2014, arXiv:1405.0312

arXiv 2014
[8]

A Review and Implementation of Object Detection Models and Optimizations for Real-time Medical Mask Detection during the COVID-19 Pandemic

Gogou, LC.; Koutsomitropoulos, D.A. A Review and Implementation of Object Detection Models and Optimizations for Real-time Medical Mask Detection during the COVID-19 Pandemic. In Proceedings of the 2022 International Conference on INnovations in Intelligent SysTems and Applications (INISTA), Biarritz, France, 8-12 August 2022; pp. 1-6. [CrossRef ]

work page 2022

Show all 37 references

[9]

Improving Vehicle Detection in Challenging Datasets: YOLOv5s and Frozen Layers Analysis

Rafi, A.N.Y.; Yusuf, M. Improving Vehicle Detection in Challenging Datasets: YOLOv5s and Frozen Layers Analysis. Int. J. Inform. Comput. 2023, 5, 31-45. [CrossRef]

2023
[10]

AutoFreeze: Automatically Freezing Model Blocks to Accelerate Fine-tuning

Liu, Y.; Agarwal, S.; Venkataraman, S. AutoFreeze: Automatically Freezing Model Blocks to Accelerate Fine-tuning. arXiv 2021, arXiv:2102.01386. [CrossRef]

2021 arXiv
[11]

PipeTransformer: Automated Elastic Pipelining for Distributed Training of Transformers

He, C.; Li, S.; Soltanolkotabi, M.; Avestimehr, S. PipeTransformer: Automated Elastic Pipelining for Distributed Training of Transformers. arXiv 2021, arXiv:2102.03161. [CrossRef]

2021 arXiv
[12]

Ultralytics YOLO

Jocher, G.; Chaurasia, A.; Qiu, J. Ultralytics YOLO. 2023. Available online: https:/ / github.com/ultralytics/ultralytics (accessed on 27 July 2025)

2023
[13]

YOLOvlO: Real-Time End-to-End Object Detection

Wang, A.; Chen, H.; Liu, L.; Chen, K.; Lin, Z.; Han, J.; Ding, G. YOLOvlO: Real-Time End-to-End Object Detection. arXiv 2024, arXiv:2405.14458. [CrossRef]

2024 arXiv
[14]

Efficient DNN Training with Knowledge-Guided Layer Freezing

Wang, Y.; Sun, D.; Chen, K.; Lai, F.; Chowdhury, M. Efficient DNN Training with Knowledge-Guided Layer Freezing. arXiv 2022, arXiv:2201.06227

2022 arXiv
[15]

Training Acceleration Method Based on Parameter Freezing

Tang, H.; Chen, J.; Zhang, W.; Guo, Z. Training Acceleration Method Based on Parameter Freezing. Electronics 2024, 13, 2140. [CrossRef]

2024
[16]

FreezeNet: Full Performance by Reduced Storage Costs

Wimmer, P.; Mehnert, J.; Condurache, A. FreezeNet: Full Performance by Reduced Storage Costs. arXiv 2020, arXiv:2011.14087. [CrossRef] Mathematics 2025, 13, 2539 31 of 31

2020 arXiv
[17]

Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask

Zhou, H.; Lan, J.; Liu, R.; Yasinski, J. Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask. arXiv 2020, arXiv:1905.01067. [CrossRef]

2020 arXiv
[18]

Features Exploitation of YOLOv5-Based Freeze Backbone for Perfor mance Improvement of UAV Object Detection

Qomariyati, L.N.; Jannah, N.; Wibowo, S.A.; Siadari, T.S. Features Exploitation of YOLOv5-Based Freeze Backbone for Perfor mance Improvement of UAV Object Detection. Int. J. Fuzzy Log. Intell. Syst. 2024, 24, 194-202. [CrossRef]

2024
[19]

Status Recognition Using Pre-Trained YOLOv5 for Sustainable Human-Robot Collaboration (HRC) System in Mold Assembly

Liau, Y.Y.; Ryu, K. Status Recognition Using Pre-Trained YOLOv5 for Sustainable Human-Robot Collaboration (HRC) System in Mold Assembly. Sustainability 2021, 13, 12044. [CrossRef]

2021
[20]

Application of Three Transfer Learning Methods Based on YOLOv8 for Object Recognition in Architectural Drawings

Feng, Y.; Sumiyoshi, D. Application of Three Transfer Learning Methods Based on YOLOv8 for Object Recognition in Architectural Drawings. In Proceedings of the ASim Conference 2024: 5th Asia Conference of IBPSA, Osaka, Japan, 8--10 November 2024; Volume 5, pp. 443--450. [CrossRef]

2024
[21]

Fine-Tuning Without Forgetting: Adaptation of YOLOv8 Preserves COCO Performance

Gandhi, V.; Gandhi, S. Fine-Tuning Without Forgetting: Adaptation of YOLOv8 Preserves COCO Performance. ar Xiv 2025, arXiv:2505.01016. [CrossRef]

2025 arXiv
[22]

CSPNet: A New Backbone that can Enhance Learning Capability of CNN

Wang, C.Y.; Liao, H.Y.M.; Yeh, I.H.; Wu, Y.H.; Chen, P.Y.; Hsieh, J.W. CSPNet: A New Backbone that can Enhance Learning Capability of CNN. arXiv 2019, arXiv:1911.11929. [CrossRef]

2019 arXiv
[23]

YOLOv4: Optimal Speed and Accuracy of Object Detection

Bochkovskiy, A.; Wang, C.Y.; Liao, H.Y.M. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv 2020, arXiv:2004.10934. [CrossRef]

2020 arXiv
[24]

A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOvl to YOLOv8 and YOLO-NAS

Terven, J.; Cordova-Esparza, D.M.; Romero-Gonzalez, J.A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOvl to YOLOv8 and YOLO-NAS. Mach. Learn. Knowl. Extr. 2023, 5, 1680-1716. [CrossRef]

2023
[25]

InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images

Vieira e Silva, A.L.B.; de Castro Felix, H.; Simoes, F.P.M.; Teichrieb, V.; dos Santos, M.; Santiago, H.; Sgotti, V.; Lott Neto, H. InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images. Int. J. Remote Sens. 2023, 44, 7294-7320. [CrossRef]

2023
[26]

A YOLO Annotated 15-class Ground Truth Dataset for Substation Equipment

Gomes, A. A YOLO Annotated 15-class Ground Truth Dataset for Substation Equipment. Data 2023, 8, 118. [CrossRef]

2023
[27]

VALID: A Comprehensive Virtual Aerial Image Dataset

Chen, L.; Liu, F.; Zhao, Y.; Wang, W.; Yuan, X.; Zhu, J. VALID: A Comprehensive Virtual Aerial Image Dataset. In Proceedings of the 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 31 May-15 June 2020; pp. 2009-2016. [CrossRef]

2020
[28]

Deep Learning-Based Bird's Nest Detection on Transmission Lines Using UAV Imagery

Li, J.; Yan, D.; Luan, K.; Li, Z.; Liang, H. Deep Learning-Based Bird's Nest Detection on Transmission Lines Using UAV Imagery. Appl. Sci. 2020, 10, 6147. [CrossRef]

2020
[29]

Review and analysis of synthetic dataset generation methods and techniques for application in computer vision

Paulin, G.; Ivasic-Kos, M. Review and analysis of synthetic dataset generation methods and techniques for application in computer vision. Artif. Intell. Rev. 2023, 56, 9221-9265. [CrossRef]

2023
[30]

Explore the power of synthetic data on few-shot object detection

Lin, S.; Wang, K.; Zeng, X.; Zhao, R. Explore the power of synthetic data on few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada, 18--22 June 2023; pp. 638-647

2023
[31]

Learning From Synthetic Data: Addressing Domain Shift for Semantic Segmentation

Sankaranarayanan, S.; Balaji, Y.; Jain, A.; Lim, S.N.; Chellappa, R. Learning From Synthetic Data: Addressing Domain Shift for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18-22 June 2018

2018
[32]

YOLO-V8-CAM

Shirvalkar, R. YOLO-V8-CAM. 2023. Available online: https:/ / github.com/rigvedrs/YOLO-V8-CAM (accessed on 27 July 2025)

2023
[33]

The Golden Ratio of Learning and Momentum

Jaeger, S. The Golden Ratio of Learning and Momentum. arXiv 2020, arXiv:2006.04751. [CrossRef]

2020 arXiv
[34]

On the importance of initialization and momentum in deep learning

Sutskever, I.; Martens, J.; Dahl, G.; Hinton, G. On the importance of initialization and momentum in deep learning. In Proceedings of the International Conference on Machine Learning, Atlanta, GA, USA, 16-21 June 2013

2013
[35]

The instabilities of large learning rate training: A loss landscape view

Wang, L.; Roberts, S. The instabilities of large learning rate training: A loss landscape view. arXiv 2023, arXiv:2307.11948. [CrossRef]

2023 arXiv
[36]

Identifying and attacking the saddle point problem in high-dimensional non-convex optimization

Dauphin, Y.; Pascanu, R.; Gulcehre, C.; Cho, K.; Ganguli, S.; Bengio, Y. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. arXiv 2014, arXiv:1406.2572. [CrossRef]

2014 arXiv
[37]

Visualizing the Loss Landscape of Neural Nets

Li, H.; Xu, Z.; Taylor, G.; Goldstein, T. Visualizing the Loss Landscape of Neural Nets. arXiv 2017, arXiv:1712.09913. Disclaimer/Publisher's Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) a...

2017 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Real-Time Flying Object Detection with YOLOv8

Reis, D.; Kupec, J.; Hong, J.; Daoudi, A. Real-Time Flying Object Detection with YOLOv8. arXiv 2024, arXiv:2305.09972. [CrossRef]

arXiv 2024

[2] [2]

A YOLO-Based Traffic Counting System

Lin, J.P.; Sun, M.T. A YOLO-Based Traffic Counting System. In Proceedings of the 2018 Conference on Technologies and Applications of Artificial Intelligence (TAAI), Taichung, Taiwan, 30 November-2 December 2018; pp. 82-85

work page 2018

[3] [3]

YOLO V3 + VGG16-based automatic operations monitoring and analysis in a manufacturing workshop under Industry 4.0

Yan, J.; Wang, Z. YOLO V3 + VGG16-based automatic operations monitoring and analysis in a manufacturing workshop under Industry 4.0. J. Manuf Syst. 2022, 63, 134-142. [CrossRef]

work page 2022

[4] [4]

YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss

Maji, D.; Nagori, S.; Mathew, M.; Poddar, D. YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss. arXiv 2022, arXiv:2204.06806. [CrossRef]

work page Pith review arXiv 2022

[5] [5]

A Comprehensive Survey on Transfer Learning

Zhuang, F.; Qi, Z.; Duan, K.; Xi, D.; Zhu, Y.; Zhu, H.; Xiong, H.; He, Q. A Comprehensive Survey on Transfer Learning. Proc. IEEE 2019, 109, 43-76. [CrossRef]

work page 2019

[6] [6]

How transferable are features in deep neural networks? arXiv 2014, arXiv:1411.1792

Yosinski, J.; Clune, J.; Bengio, Y.; Lipson, H. How transferable are features in deep neural networks? arXiv 2014, arXiv:1411.1792. [CrossRef]

arXiv 2014

[7] [7]

Microsoft COCO: Common Objects in Context

Lin, T.; Maire, M.; Belongie, S.J.; Bourdev, L.D.; Girshick, R.B.; Hays, J.; Perona, P.; Ramanan, D.; Dollar, P.; Zitnick, C.L. Microsoft COCO: Common Objects in Context. arXiv 2014, arXiv:1405.0312

arXiv 2014

[8] [8]

A Review and Implementation of Object Detection Models and Optimizations for Real-time Medical Mask Detection during the COVID-19 Pandemic

Gogou, LC.; Koutsomitropoulos, D.A. A Review and Implementation of Object Detection Models and Optimizations for Real-time Medical Mask Detection during the COVID-19 Pandemic. In Proceedings of the 2022 International Conference on INnovations in Intelligent SysTems and Applications (INISTA), Biarritz, France, 8-12 August 2022; pp. 1-6. [CrossRef ]

work page 2022