Pith. sign in

REVIEW 4 major objections 7 minor 33 references

Object Detection for Autonomous Driving in Chinese Rural Scenes: An Experimental Study on Real-Synthetic Data Mixing and Model Evaluation

T0 review · 4 major / 7 minor · reviewed 2026-07-30 · grok-4.5

Pith's one-line read A moderate 1:0.5 mix of real and synthetic Chinese rural road images raises object-detection accuracy over all-real training; a 1:1 mix does not.

desk verdict Useful rural-China detection bake-off with a real-only test set; the 1:0.5 win is directionally plausible but statistically thin. read the letter →

arxiv 2607.27058 v1 pith:LSPEXO5G submitted 2026-07-29 cs.CV

classification cs.CV
keywords autonomousdrivingobjectdetectionruralscenessyntheticdataYOLORT-DETRreal-syntheticmixingChineseroads
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

Autonomous driving detectors trained on urban Western datasets struggle on Chinese rural roads, where electric tricycles, low-speed vehicles, stalls, and uneven infrastructure are common and labeled data are scarce. This paper builds a 14-class detection dataset that pairs dashboard video from Weishi County, Henan, with Unreal Engine scenes of the same rural elements, then trains 13 mainstream detectors under identical settings on three mixes: all real, real:synthetic 1:0.5, and 1:1. On a held-out real-only test set, the 1:0.5 mix improves several YOLO models, with YOLO11m reaching the best mAP@0.5 of 0.758, while the 1:1 mix introduces domain shift that cancels the benefit of more images. Medium-scale models outperform smaller ones; local vehicles are detected reliably, but stalls and railings remain weak across CNN and Transformer architectures. The results give concrete guidance on how much synthetic data to add and which model families to prefer for rural perception.

What carries the argument

The hybrid real-synthetic 14-category rural dataset (Weishi County video plus Unreal Engine scenes) evaluated under three fixed real-to-synthetic ratios with identical hyperparameters across 13 detectors.

What would settle it

Repeat the same three mix ratios and model set using real rural video from a different province or season (and/or a true vehicle camera), keep the synthetic pipeline fixed, and check whether 1:0.5 still beats all-real and whether 1:1 still drops on a pure-real test set; if the inverted-U or the model ranking disappears, the central claim fails.

Watch

Extended reading notes

Core claim

Under a unified training protocol on a new 14-class Chinese rural dataset, a moderate real-to-synthetic mix (1:0.5) improves real-test detection over all-real training for several YOLO models, with YOLO11m reaching the highest mAP@0.5 of 0.758, whereas a 1:1 mix introduces domain shift that offsets data-scaling gains; architectures differ in sensitivity to that shift, and long-tail non-standard objects such as stalls and railings stay poorly detected.

Load-bearing premise

That one county's winter dashboard footage plus commercial Unreal Engine assets is representative enough of Chinese rural roads for the optimal mix ratio and model rankings to transfer to other places, seasons, and sensors.

Editorial extensions

If this is right

  • For this style of synthetic rural data, a real:synthetic ratio near 1:0.5 is a better default than pure real or 1:1.
  • Medium YOLO11 and YOLOv8 are preferable to nano/small scales and to RT-DETR when synthetic data is mixed in.
  • Once tricycles and low-speed vehicles are explicit classes in a mixed dataset, mainstream detectors can learn them reliably.
  • Stalls and elongated railings need dedicated data or architecture work; more generic synthetic volume alone does not fix them.
  • Higher parameter capacity helps model the variable appearance of rural objects under the same training recipe.

Reading between the lines

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

  • Closing the synthetic-to-real texture and lighting gap (higher-fidelity assets or generative refinement) may raise the useful mix ratio above 1:0.5 without the observed drop.
  • The same inverted-U mixing pattern is likely in other unstructured domains that rely on game-engine labels, such as construction sites or ports.
  • Switching from phone-dashboard capture to automotive cameras could shift the optimal ratio even if geography is held fixed.
  • Specialized heads or losses for elongated and highly variable objects may outperform simply adding more generic synthetic volume for stalls and railings.
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 / 7 minor

Summary. The paper constructs a 14-class real–synthetic object detection dataset for Chinese rural roads (real dashboard video from Weishi County, Henan, plus Unreal Engine scenes) and evaluates 13 detectors (YOLOv5/v8/11/26 at n/s/m scales and RT-DETR-L) under a unified protocol on three mixes: all-real (Config A), 1:0.5 (B), and 1:1 (C). A fixed 517-image real-only test set is used. The central empirical claim is that moderate synthetic injection (1:0.5) improves real-test mAP for several YOLO models—YOLO11m reaches 0.758 mAP@0.5—while 1:1 introduces domain shift that offsets scaling gains; models remain weak on long-tail classes (stalls, railings). The work positions itself as guidance for model choice and synthetic-data strategy in rural AD perception.

Significance. Rural Chinese traffic is genuinely under-covered by standard AD benchmarks, and a 14-class taxonomy that includes tricycles, LSVs, stalls, and railings is a useful contribution if the data are released. The experimental design has clear strengths: a real-only held-out test set, fixed hyperparameters across 13 models, three explicit mix ratios, multi-scale YOLO plus RT-DETR, and per-class breakdowns under the best config. If the inverted-U mixing result and architecture-sensitivity findings hold under proper variance control, the paper would provide actionable, if incremental, empirical guidance for practitioners. The contribution is primarily an application-oriented bake-off and dataset construction study rather than a methodological or theoretical advance.

major comments (4)
  1. [§IV, Tables V–VIII, Fig. 4–5] Central claim (Abstract, §IV, Tables V–VIII, Fig. 4–5): the reported +0.025 mAP@0.5 lift for YOLO11m (0.733→0.758) and the inverted-U across configs rest on a single training run per cell with no seeds, error bars, or statistical tests. Under default YOLO training, seed-to-seed mAP variation of this magnitude is common. Without multi-seed means/std or at least repeated runs on the headline models, it is not established that the 1:0.5 advantage is a reproducible mixing effect rather than run noise. This is load-bearing for every claim that “moderate injection effectively enhances” performance and that 1:1 “offsets” gains.
  2. [§II.C Table II, §IV] Table II / Config design: Configs B and C keep the same 4,200 real images and only add synthetic data (total 6,300 and 8,400). There is no matched all-real control with equal total N, no synthetic-only arm, and no real-data subsample that matches B/C epoch exposure. The inverted-U therefore confounds mix ratio with total data volume and effective gradient steps over a larger set. The domain-shift interpretation of the 1:1 drop is plausible but not isolated from this confound. A minimal fix is either (i) multi-seed runs plus a volume-matched real baseline, or (ii) explicit ablation that holds total N fixed while varying the real:synthetic fraction.
  3. [§II.A, Abstract, §V] §II.A and §V generalization language: real data are from one county, one month (Feb 2026), ~85 km, smartphone 1280×720 dashboard capture; synthetic scenes use commercial Fab/Sketchfab assets. The manuscript repeatedly frames results as evidence for “Chinese rural roads” and “practical deployment in rural areas.” Representativeness is asserted (median socio-economic indicators, Central Plains typology) but not tested (no second region, season, or sensor). The within-dataset ranking can still be published, but claims that the optimal 1:0.5 ratio and model order transfer should be sharply scoped to this collection setup, or supported by at least a small external real split.
  4. [§IV, Fig. 5, Table VIII] §IV architecture-sensitivity narrative (esp. RT-DETR-L drop 0.736→0.692 and YOLO26 flatness): explanations invoke global attention amplifying texture/illumination mismatch and YOLO26’s one-to-one head / progressive loss reducing distribution sensitivity. These are post-hoc and untested—no feature CKA, domain-gap metrics (e.g., FID/MMD between real and synthetic), or controlled ablations of loss/head. Given single-run results, the strong causal language about Transformer vs CNN sensitivity should be toned down or backed by measurement.
minor comments (7)
  1. [Table III] Table III lists only YOLO parameter counts; RT-DETR-L (42M) is mentioned in text but omitted from the table. Include it for completeness.
  2. [Fig. 4] Fig. 4 is hard to read: many overlapping series and small markers. Consider separate panels per family or a table-first presentation with the figure as a supplement-style overview.
  3. [§II.A] Annotation protocol (§II.A): two-round cross-validation is good; report inter-annotator agreement (e.g., box-level IoU or category confusion) if available.
  4. [§III.C–D, Table IV] Hyperparameters (Table IV): state batch sizes per platform and whether they differed between Config A (RTX 4060) and B/C (RTX 4090), since effective batch size can affect SGD dynamics even if final metrics are compared on the same test set.
  5. [§I, §III.A] Naming: “YOLO26” will confuse readers familiar with the public YOLO lineage through v9/v10/v11; briefly cite the exact release/weights source used.
  6. [§II, references] Minor prose/typos: “DatasetConstruction” spacing in §II heading; “V oronin” / “Laotoule” consistency; ensure mAP vs mAP@0.5 is uniform in figure captions.
  7. [end matter] Data and code availability are not stated. For an empirical bake-off whose value is the dataset and protocol, a release plan (or justified restriction) should be explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical bake-off of detectors and mixing ratios on a held-out real test set; results are measured, not derived from inputs by construction.

full rationale

The paper’s load-bearing claims are experimental comparisons (mAP@0.5 / F1 of 13 detectors under all-real, 1:0.5, and 1:1 real:synthetic training configs on a fixed 517-image real-only test set). Mixing ratios are controlled factors (Table II), not quantities fitted from the reported metric. Test images contain no synthetic data and are held out from training. There is no self-definitional loop, no fitted parameter renamed as a prediction, no uniqueness theorem or ansatz imported via author self-citation, and no renaming of a known closed-form result. Citations are to external datasets, detectors, and prior synthetic-data studies. Methodological limits (single-run variance, fixed real-image count, single-county collection) affect statistical robustness and transfer, not circularity of the derivation chain. Score 0 is appropriate.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

This is an empirical systems paper. Load-bearing premises are domain assumptions about geographic representativeness and synthetic-to-real transfer, plus standard detection-training practice—not a formal derivation. Few numeric free parameters are fitted to support a theory; mix ratios are design choices. No new physical entities are postulated.

free parameters (3)
  • real:synthetic mix ratios {1:0, 1:0.5, 1:1} = 1:0.5 reported best among three
    Hand-chosen experimental grid; the headline 'optimal 1:0.5' is selected from this discrete set rather than derived. Finer grids or other generators might move the peak.
  • default YOLO/RT-DETR training hyperparameters (lr0, mosaic, loss weights, 100 epochs, 640 input) = Table IV defaults
    Taken as framework defaults without per-model or per-config tuning; performance rankings partly reflect these fixed choices.
  • train/val 7:3 split and 517-image real test holdout = 4200 real train-pool; 517 test
    Single fixed partition; no cross-validation. Test selection 'carefully selected to include all 14 categories' is not fully specified.
assumptions (5)
  • domain assumption Weishi County roads and traffic mix are representative enough of broader Chinese rural conditions for model and data-strategy conclusions to transfer.
    Invoked in §II.A (site selection) and used to frame abstract/conclusion claims about 'Chinese rural' deployment.
  • domain assumption Unreal Engine 5.7 renders with commercial assets and EasySynth labels share enough low- and mid-level statistics with real dashboard video that mixing them is a valid augmentation rather than pure noise.
    Foundation of Configs B/C; domain gap is discussed post hoc when 1:1 fails but not measured a priori.
  • domain assumption mAP@0.5 and F1 on a 517-image real test set are adequate proxies for rural AV perception utility.
    Sole reported metrics; no latency-on-target-hardware, calibration, or closed-loop driving metrics.
  • standard math Standard COCO-pretrained YOLO/RT-DETR fine-tuning with SGD/AdamW and the listed augmentations is a fair comparison protocol.
    §III unified protocol; conventional in the detection literature.
  • ad hoc to paper Smartphone 1280×720 dashboard capture approximates vehicle front-camera optics for this study.
    §II.A acquisition setup; no validation against automotive cameras.
invented entities (2)
  • 14-category Chinese rural object taxonomy (including tricycle, LSV, stall, billboard, railing, etc.)
    purpose: Define label space that covers rural long-tail participants missing from KITTI/BDD/nuScenes-style sets.
    Category list is a design choice of the dataset, not a discovered natural kind; evaluation metrics inherit this ontology.
  • Weishi real–Unreal hybrid detection dataset (Configs A/B/C)
    purpose: Enable controlled real/synthetic ratio experiments for rural detection.
    New resource constructed for this paper; not independently validated by external benchmarks or third-party releases in the text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Object Detection for Autonomous Driving in Chinese Rural Scenes: An Experimental Study on Real-Synthetic Data Mixing and Model Evaluation." pith.science (2026). https://pith.science/paper/LSPEXO5G

@misc{pith2026260727058,
  author       = {Pith},
  title        = {Pith review of: Object Detection for Autonomous Driving in Chinese Rural Scenes: An Experimental Study on Real-Synthetic Data Mixing and Model Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSPEXO5G}},
  note         = {Machine review of arXiv:2607.27058}
}
read the original abstract

Currently, autonomous driving object detection models face significant data scarcity and generalization challenges when navigating complex Chinese rural traffic scenarios. To address these limitations, we propose a novel real-synthetic mixed object detection dataset tailored specifically for Chinese rural roads and systematically evaluate the performance of 13 mainstream detectors under different real-to-synthetic data ratios, thereby providing empirical evidence for model selection and data strategy design in rural autonomous driving scenarios. Our dataset combines real-world images captured in Weishi County, Henan Province, with parameterized virtual scenes generated via Unreal Engine. To accurately reflect the unique realities of rural traffic, we define a comprehensive 14-category object system encompassing region-specific elements such as electric tricycles, low-speed vehicles (LSVs), and roadside stalls. Under a unified training protocol, we systematically evaluate 13 mainstream detectors -- spanning the YOLOv5, YOLOv8, YOLO11, and YOLO26 series, as well as RT-DETR-L -- across three data configurations: an all-real baseline, a 1:0.5 real-to-virtual mix, and a 1:1 mix. Experimental results demonstrate that a moderate injection of synthetic data (1:0.5 ratio) effectively enhances detection performance, with YOLO11m achieving the highest mAP@0.5 of 0.758. However, a higher proportion of synthetic data (1:1) introduces domain shifts that offset the benefits of data scaling. While most models reliably identify distinct local vehicles, significant perceptual bottlenecks remain for long-tail, non-standard objects like stalls and railings. This research provides crucial empirical evidence and novel insights for model selection and synthetic data strategies, facilitating the practical deployment of autonomous driving perception systems in rural areas.

Figures

Figures reproduced from arXiv: 2607.27058 by the authors.

Figure 1
Figure 1. Satellite overview of Weishi County, Henan Province. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) County road, (b) Market town center, (c) Village street. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Virtual scenes of different road sections and their corresponding [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: mAP@0.5 and F1-score of different YOLO model sizes and config [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: mAP@0.5 of medium-scale YOLO models and RT-DETR-L under [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Per-class mAP@0.5 of different models under Configuration B. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 6 linked inside Pith

  1. [1]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In2012 IEEE conference on computer vision and pattern recognition, pages 3354–

  2. [2]

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

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016

  3. [3]

    Yolo9000: better, faster, stronger

    Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263–7271, 2017

  4. [4]

    Yolov3: An incremental improvement

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. TABLE IX PER-CLASS MAP@0.5OF DIFFERENT MODELS UNDERCONFIGURATIONB. stall LSV tricycle scooter bin person car truck billboard railing pole sign tree street lamp YOLO11m 0.56 0.863 0.879 0.782 0.853 0.725 0.947 0.657 0.839 0.47 0.798 0.659 0.775 0.812 Y...

  5. [5]

    Yolov4: Optimal speed and accuracy of object detection.arXiv preprint arXiv:2004.10934, 2020

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection.arXiv preprint arXiv:2004.10934, 2020

  6. [6]

    Yolov5 https://github

    Glenn Jocher, K Nishimura, T Minerva, and R Vilari ˜no. Yolov5 https://github. com/ultralytics/yolov5.Accessed March, 7:2021, 2020

  7. [7]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision, pages 213–229. Springer, 2020

  8. [8]

    Detrs beat yolos on real- time object detection

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real- time object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16965–16974, 2024

Show all 33 references
  1. [9]

    Evaluating yolo architectures: Implications for real-time vehicle detection in urban environments of bangladesh.arXiv preprint arXiv:2509.05652, 2025

    Ha Meem Hossain, Pritam Nath, Mahitun Nesa Mahi, Imtiaz Uddin, Ishrat Jahan Eiste, Syed Nasibur Rahman Ratul, Md Naim Uddin Mozumdar, Asif Mohammed Saad, and MD Hossain. Evaluating yolo architectures: Implications for real-time vehicle detection in urban environments of bangla...

  2. [10]

    Domain generalization in autonomous driving: Evaluating yolov8s, rt-detr, and yolo-nas with the road-almaty dataset.arXiv preprint arXiv:2412.12349, 2024

    Madiyar Alimov and Temirlan Meiramkhanov. Domain generalization in autonomous driving: Evaluating yolov8s, rt-detr, and yolo-nas with the road-almaty dataset.arXiv preprint arXiv:2412.12349, 2024

  3. [11]

    First qualitative observations on deep learning vision model yolo and detr for automated driving in austria.arXiv preprint arXiv:2312.12314, 2023

    Stefan Schoder. First qualitative observations on deep learning vision model yolo and detr for automated driving in austria.arXiv preprint arXiv:2312.12314, 2023

  4. [12]

    Revisiting unreasonable effectiveness of data in deep learning era

    Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843–852, 2017

  5. [13]

    Vision meets robotics: The kitti dataset.The international journal of robotics research, 32(11):1231–1237, 2013

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset.The international journal of robotics research, 32(11):1231–1237, 2013

  6. [14]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pag...

  7. [15]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern re...

  8. [16]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. InProceedings of the IEEE/CVF conference on computer vis...

  9. [17]

    Semantic segmentation network for unstructured rural roads based on improved sppm and fused multiscale features.Applied Sciences, 14(19):8739, 2024

    Xinyu Cao, Yongqiang Tian, Zhixin Yao, Yunjie Zhao, and Taihong Zhang. Semantic segmentation network for unstructured rural roads based on improved sppm and fused multiscale features.Applied Sciences, 14(19):8739, 2024

  10. [18]

    Construction and enhancement of a rural road instance segmentation dataset based on an improved stylegan2-ada

    Zhixin Yao, Renna Xi, Taihong Zhang, Yunjie Zhao, Yongqiang Tian, and Wenjing Hou. Construction and enhancement of a rural road instance segmentation dataset based on an improved stylegan2-ada. Sensors, 25(8):2477, 2025

  11. [19]

    D 2-city: a large- scale dashcam video dataset of diverse traffic scenarios.arXiv preprint arXiv:1904.01975, 2019

    Zhengping Che, Guangyu Li, Tracy Li, Bo Jiang, Xuefeng Shi, Xinsheng Zhang, Ying Lu, Guobin Wu, Yan Liu, and Jieping Ye. D 2-city: a large- scale dashcam video dataset of diverse traffic scenarios.arXiv preprint arXiv:1904.01975, 2019

  12. [20]

    M4sfwd: A multi-faceted synthetic dataset for remote sensing forest wildfires detection.Expert Systems with Applications, 248:123489, 2024

    Guanbo Wang, Haiyan Li, Peng Li, Xun Lang, Yanling Feng, Zhaisehng Ding, and Shidong Xie. M4sfwd: A multi-faceted synthetic dataset for remote sensing forest wildfires detection.Expert Systems with Applications, 248:123489, 2024

  13. [21]

    Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task

    Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

  14. [22]

    Detection and tracking meet drones chal- lenge.IEEE transactions on pattern analysis and machine intelligence, 44(11):7380–7399, 2021

    Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones chal- lenge.IEEE transactions on pattern analysis and machine intelligence, 44(11):7380–7399, 2021

  15. [23]

    Enhancing object detection accuracy in autonomous vehicles using synthetic data

    Sergei V oronin, Abubakar Siddique, and Muhammad Iqbal. Enhancing object detection accuracy in autonomous vehicles using synthetic data. arXiv preprint arXiv:2411.15602, 2024

  16. [24]

    Experimental results on synthetic data generation in unreal engine 5 for real-world object detection

    Alexandru Damian, Claudiu Filip, Anamaria Nistor, Irina Petrariu, C˘at˘alin Mariuc, and Valentin Stratan. Experimental results on synthetic data generation in unreal engine 5 for real-world object detection. In 2023 17th international conference on engineering of modern electr...

  17. [25]

    Experimental study on using synthetic images as a portion of training dataset for object recognition in construction site.Buildings, 14(5):1454, 2024

    Jaemin Kim, Ingook Wang, and Jungho Yu. Experimental study on using synthetic images as a portion of training dataset for object recognition in construction site.Buildings, 14(5):1454, 2024

  18. [26]

    Optimizing object detection for maritime search and rescue: Progressive fine-tuning of yolov9 with real and synthetic data

    Luciano Netto de Lima, Fabio Augusto de Alcantara Andrade, Youcef Djenouri, Carlos Pfeiffer, and Marcos Moura. Optimizing object detection for maritime search and rescue: Progressive fine-tuning of yolov9 with real and synthetic data. InICAART (3), pages 209–216, 2025

  19. [27]

    Sim2real diffusion: Leveraging foundation vision language models for adaptive automated driving.IEEE Robotics and Automation Letters, 11(1):177– 184, 2025

    Chinmay Samak, Tanmay Samak, Bing Li, and Venkat Krovi. Sim2real diffusion: Leveraging foundation vision language models for adaptive automated driving.IEEE Robotics and Automation Letters, 11(1):177– 184, 2025

  20. [28]

    Synth it like kitti: Synthetic data generation for object detection in driving scenarios

    Richard Marcus, Christian V ogel, Inga Jatzkowski, Niklas Knoop, and Marc Stamminger. Synth it like kitti: Synthetic data generation for object detection in driving scenarios. InInternational Conference on Robotics, Computer Vision and Intelligent Systems, pages 414–432. Sprin...

  21. [29]

    Synthetic data for video surveillance applications of computer vision: A review.Interna- tional Journal of Computer Vision, 132(10):4473–4509, 2024

    Rita Delussu, Lorenzo Putzu, and Giorgio Fumera. Synthetic data for video surveillance applications of computer vision: A review.Interna- tional Journal of Computer Vision, 132(10):4473–4509, 2024

  22. [30]

    Object detector differences when using synthetic and real training data: Mg ljungqvist et al.SN computer science, 4(3):302, 2023

    Martin Georg Ljungqvist, Otto Nordander, Markus Skans, Arvid Mild- ner, Tony Liu, and Pierre Nugues. Object detector differences when using synthetic and real training data: Mg ljungqvist et al.SN computer science, 4(3):302, 2023

  23. [31]

    Pcgod: Enhancing object detection with synthetic data for scarce and sensitive computer vision tasks.IEEE Access, 2025

    Walid Remmas, Martin Lints, and Jaak Joonas Uudm ¨ae. Pcgod: Enhancing object detection with synthetic data for scarce and sensitive computer vision tasks.IEEE Access, 2025

  24. [32]

    V olucapture: Multi- view synthetic data capture in unreal engine

    Sandeep Bangalore Venkatesh and Guan-Ming Su. V olucapture: Multi- view synthetic data capture in unreal engine. In2025 IEEE 8th International Conference on Multimedia Information Processing and Retrieval (MIPR), pages 551–554. IEEE, 2025

  25. [33]

    Syndra: Synthetic dataset for railway applications

    Gianluca D’Amico, Federico Nesti, Giulio Rossolini, Mauro Marinoni, Salvatore Sabina, and Giorgio Buttazzo. Syndra: Synthetic dataset for railway applications. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3437–3446. IEEE, 2025

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.