Pith. sign in

REVIEW 3 major objections 5 minor 43 references

TRUDI and TITUS: A Multi-Perspective Dataset and A Three-Stage Recognition System for Transportation Unit Identification

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper presents TRUDI, the first publicly available dataset for identifying containers, tank containers, and trailers from aerial and ground images, and TITUS, a three-stage pipeline that reads and validates their 11-character ISO 6346…

desk verdict TRUDI is a genuinely new and useful dataset for port logistics vision, but TITUS is an off-the-shelf pipeline whose abstract overclaims reliability against its own end-to-end numbers. read the letter →

arxiv 2508.02372 v1 pith:ORDODTZN submitted 2025-08-04 cs.CV

Emre Gülsoylu , André Kelm , Lennart Bengtson , Matthias Hirsch , Christian Wilms , Tim Rolff , Janick Edinger , Simone Frintrop This is my paper
classification cs.CV
keywords transportationunitidentificationISO6346containercodesbenchmarkdatasetinstancesegmentationscenetextdetectionrecognitionaerialimageryportlogistics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the absence of a public benchmark is the main obstacle to reliable identification of transportation units in busy ports, and that a dataset collected from moving cameras—drones, terminal trucks, and reach stackers—can fill that gap. It releases TRUDI, 35,034 annotated instances across five classes (container, tank container, trailer, ID text, logo) from 733 images captured over 18 months in different countries, lighting conditions, weather, and altitudes. On this dataset it introduces TITUS, a three-stage pipeline that segments each unit, crops the segment to find the ISO 6346 ID text, and then recognizes and validates the 11-character code. The evaluation shows the pipeline identifies units across perspectives, with aerial imagery easier than ground imagery, and establishes the first common yardstick against which future methods can be compared.

What carries the argument

The load-bearing object is the TRUDI dataset itself: 733 images with 35,034 mask instances labeled as container, tank container, trailer, ID text, or logo, split 60/20/20 with stratification by brightness, sharpness, and contrast so that no subset is dominated by one image-quality regime. The carrying mechanism of the method is the three-stage crop cascade: Mask R-CNN segments each transportation unit; DBNet++ searches for ID text only inside the unit's cutout, shrinking the search space and blocking distractors from background units; RobustScanner reads each detected text crop and the output is checked against the ISO 6346 format and associated with its unit. This cascade is what lets the system work from mobile cameras rather than fixed gates.

What would settle it

Take a random sample of TRUDI ID-text instances, have an independent annotator re-read the codes from the images, and compare character-by-character with the released labels; if agreement is far below the recognition models' character recall (about 0.8-0.9), the ground truth rather than the pipeline is the limiting factor. A stronger check is to match recognized IDs against a cooperating port's operational records for the same physical units and see whether the 11-character codes agree.

Watch

Extended reading notes

Core claim

The central claim is that multi-perspective, real-condition transportation-unit identification is now a tractable benchmarking problem rather than a set of port-specific gate systems. The paper's discovery is a dataset-plus-pipeline pairing: TRUDI supplies the missing public data, and TITUS shows that a stagewise design—instance segmentation with Mask R-CNN (Swin backbone), text detection with DBNet++ on unit crops, and text recognition with RobustScanner plus ISO 6346 validation—can read IDs across both aerial and ground views. The reported end-to-end performance on the combined test set is precision 0.39, recall 0.27, F1 0.32, accuracy 0.19, with ground perspective the hardest because of motion blur, vertical code orientation, and occlusion. These numbers are offered as a baseline, not as a finished system, and the paper argues that the dataset's diversity is what makes the benchmark useful.

Load-bearing premise

The load-bearing premise is that the TRUDI annotations, especially the transcribed ID texts, are correct and complete; the paper describes guidelines and review but gives no inter-annotator agreement or independent check against port records, so if the labels contain systematic errors every stage and end-to-end number in the paper is affected.

Editorial extensions

If this is right

  • TRUDI gives the field a shared testbed, so container-code systems that were previously compared only on private single-port data can now be evaluated on the same aerial and ground images.
  • Because combined-perspective fine-tuning improves text detection and recognition, future systems should not specialize to one viewpoint; aerial and ground data reinforce each other.
  • The three stages are modular, so each can be upgraded independently; for example, replacing any stage's model should produce a directly measurable change in end-to-end accuracy.
  • The pipeline's per-unit association of mask and validated ID means a georeferenced image can localize a physical container or trailer, supporting inventory tracking and digital-twin applications.
  • Ground-level mobile imagery is identified as the bottleneck, so research effort on motion blur, vertical text, and occlusion should pay off more than further work on aerial views.

Reading between the lines

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

  • Beyond the paper: because operational cameras record video, temporal fusion is a natural extension—tracking a unit across frames and accumulating ID hypotheses could turn the per-frame accuracy of about 0.19 into substantially higher per-unit identification rates; the paper mentions temporal continuity but does not implement it.
  • Beyond the paper: the dataset's brightness, sharpness, and contrast stratification could be used to quantify which image-quality factors drive failures, a breakdown the paper does not report.
  • Beyond the paper: the four-letter owner-code prefix in ISO 6346 is registered by a central body, so recognized IDs could be checked against the owner-code registry as an automated external consistency signal for both recognition and annotation errors.
  • Beyond the paper: adding an explicit unreadable-marking label would separate cases where the ID is genuinely absent or illegible from cases where the pipeline fails, making benchmark numbers easier to interpret.
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

3 major / 5 minor

Summary. The paper introduces TRUDI, a publicly released dataset of 35,034 annotated instances (container, tank container, trailer, ID text, logo) from aerial and ground cameras in operational port settings, spanning varied lighting and weather conditions. It also proposes TITUS, a three-stage recognition pipeline that segments transport units, detects their ID text regions, and recognizes/validates the 11-digit ISO-alphanumeric IDs. The authors evaluate each stage (Mask R-CNN/RTMDet for segmentation, DBNet++/DBNet/PANet for text detection, RobustScanner/SVTR/SAR for recognition) and report an end-to-end evaluation on aerial, ground, and combined subsets.

Significance. The dataset is a potentially valuable contribution: public benchmarks for port-logistics TU identification are scarce, and the authors provide a large, multi-perspective, multi-condition collection with instance masks and text annotations, along with code and staged baselines. If the data are released with the claimed annotations, the resource could enable reproducible comparisons in a niche but practically important application area. The three-stage pipeline is a reasonable baseline structure. However, the paper's central reliability claim is not supported by the reported numbers, and the benchmark's validity depends on unresolved questions about temporal leakage and annotation quality.

major comments (3)
  1. [Abstract; Table 6] The abstract states that TITUS "reliably identifies TUs from a range of camera perspectives and in varying lighting and weather conditions," but the end-to-end results in Table 6 contradict this wording: accuracy is 0.22 (aerial), 0.12 (ground), and 0.19 (combined), with F1 scores of 0.36, 0.22, and 0.32, respectively. A system that fails on roughly 80% of test instances is not "reliable" under any standard use of the term. The claim should be softened to describe TITUS as a first baseline, or the evaluation should be reframed around what "reliable" means for the intended video-stream application (e.g., temporal aggregation across frames). As written, the headline claim is contradicted by the paper's own evidence.
  2. [Section 3 (Dataset split)] The train/validation/test split is stratified by brightness, sharpness, and contrast bins, not by recording session, timestamp, geolocation, or TU identity. The data were captured with vehicle-mounted and UAV cameras during active port operations, which naturally produces video-like sequences with strong temporal correlation. The stated redundancy guard—excluding only frames "showing the same objects from the identical viewpoint"—does not prevent near-duplicate frames from the same continuous capture appearing in both training and test subsets. This temporal leakage would inflate all stage-wise and end-to-end metrics in Tables 3–6. The split should be performed on recording-session or TU-instance boundaries, or at minimum the authors should quantify frame-level similarity between train and test to rule out leakage.
  3. [Section 3 (Annotation quality)] The paper asserts that "comprehensive guidelines" were used and that annotators and reviewers checked quality, but it provides no inter-annotator agreement, no sample-based error analysis, and no external validation against port records or ground-truth IDs. Because every experimental result in Tables 3–6 and the benchmark's downstream utility rest on the correctness and completeness of the 35,034 labels—especially the ID-text annotations, which are only partially associated with instance masks—the absence of a quantitative label-quality measure is a load-bearing gap. At minimum, a reported IAA statistic (e.g., IoU agreement on masks and edit distance on ID strings) or a failure-analysis of ID-level mistakes is needed.
minor comments (5)
  1. [Section 5 (Text Recognition)] Table 5's caption says "Since the ID codes are always uppercase and do not include symbols, the case and symbols are ignored." However, the preceding text notes that the recognized "word" may also be a trailer's registration plate, which can contain characters not covered by the ISO6346 scheme; the evaluation protocol for such non-ISO strings should be stated explicitly.
  2. [References; Section 5 (Text Recognition)] RobustScanner is cited as [29] in Section 5, but reference [29] is a 2004 paper on a different container-code recognition system; the correct reference for RobustScanner is [39] (Yue et al.).
  3. [Figure 2] The caption reads "ISO6436" but the standard referenced in the text is ISO6346; correct the typo.
  4. [Throughout] There are recurring spacing and capitalization inconsistencies, e.g., "UA Vs" instead of "UAVs" and the truncated equation in Table 2's column header ("Large (96 2 <area)") where the square is missing; these should be cleaned up.
  5. [Section 6 (Conclusion)] The conclusion that models trained on TRUDI "can handle real-world scenarios effectively" is an overclaim given the moderate stage-wise accuracies and low end-to-end numbers; suggest rephrasing to "can be evaluated on realistic conditions."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TRUDI/TITUS is an empirical dataset and benchmark paper; reported numbers are held-out measurements rather than quantities forced by construction.

full rationale

The paper contains no derivation chain of the kind that can be circular: it introduces a dataset, fine-tunes standard models on a training split, and reports held-out metrics. The models are pre-trained on external data (COCO, SynthText, ICDAR15) and then fine-tuned on TRUDI; evaluation is performed on a dedicated test split. The statement 'We selected the models based on their performance on the TRUDI dataset' describes a normal model-selection procedure using the validation split, not a fitted parameter renamed as a prediction. The abstract's reliability claim rests on Tables 3 through 6, which are empirical measurements on held-out images, not results equal by construction to the training labels or to the authors' prior work. Reference [34] is a same-group prior paper, but it is used only as related-work support for the difficulty of mobile-camera settings; no uniqueness theorem or central premise is imported from it. Annotation-quality and temporal-leakage concerns are legitimate correctness risks for the benchmark, but they are not definitional circularity: they do not make any reported number equivalent to its input by construction. No equations are present from which a self-definitional reduction could be exhibited, so the honest finding is no significant circularity, with a score of 0.

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

No new physical or mathematical entities are introduced. TRUDI is a dataset and TITUS is a software pipeline, not postulates needing independent evidence. The main unstated costs are the quality of the annotations and the representativeness of the image collection, both of which are treated as assumptions.

free parameters (3)
  • Train/validation/test split proportions = 60/20/20
    Chosen by hand; all reported metrics depend on this split and no sensitivity analysis is given.
  • Stratification bins for brightness, sharpness, contrast = uniform bin range, details in supplementary
    The split was stratified using these hand-picked image quality features, which influences which images appear in each subset.
  • Fine-tuning hyperparameters = not reported
    Learning rates, epochs, batch sizes, and augmentation are omitted, so the numerical results cannot be reproduced without guessing.
assumptions (3)
  • domain assumption The ground-truth labels in TRUDI are accurate and complete.
    Section 3 describes annotation guidelines and review but reports no inter-annotator agreement or external audit against port records. All stage and end-to-end scores assume these labels are correct.
  • domain assumption Images collected over 18 months are representative of the claimed weather, lighting, and altitude diversity.
    The paper asserts diversity but does not report the distribution of these conditions or a metadata table, so representativeness is an assumption.
  • domain assumption Pretraining on COCO, SynthText, and ICDAR15 transfers positively to TRUDI.
    All three stages start from external pretrained weights; the paper does not compare with from-scratch training, so the reported performance depends on this transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TRUDI and TITUS: A Multi-Perspective Dataset and A Three-Stage Recognition System for Transportation Unit Identification." pith.science (2026). https://pith.science/paper/ORDODTZN

@misc{pith2026250802372,
  author       = {Pith},
  title        = {Pith review of: TRUDI and TITUS: A Multi-Perspective Dataset and A Three-Stage Recognition System for Transportation Unit Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ORDODTZN}},
  note         = {Machine review of arXiv:2508.02372}
}
read the original abstract

Identifying transportation units (TUs) is essential for improving the efficiency of port logistics. However, progress in this field has been hindered by the lack of publicly available benchmark datasets that capture the diversity and dynamics of real-world port environments. To address this gap, we present the TRUDI dataset-a comprehensive collection comprising 35,034 annotated instances across five categories: container, tank container, trailer, ID text, and logo. The images were captured at operational ports using both ground-based and aerial cameras, under a wide variety of lighting and weather conditions. For the identification of TUs-which involves reading the 11-digit alphanumeric ID typically painted on each unit-we introduce TITUS, a dedicated pipeline that operates in three stages: (1) segmenting the TU instances, (2) detecting the location of the ID text, and (3) recognising and validating the extracted ID. Unlike alternative systems, which often require similar scenes, specific camera angles or gate setups, our evaluation demonstrates that TITUS reliably identifies TUs from a range of camera perspectives and in varying lighting and weather conditions. By making the TRUDI dataset publicly available, we provide a robust benchmark that enables the development and comparison of new approaches. This contribution supports digital transformation efforts in multipurpose ports and helps to increase the efficiency of entire logistics chains.

Figures

Figures reproduced from arXiv: 2508.02372 by the authors.

Figure 1
Figure 1. Sample images from the TRUDI dataset showing ground and aerial perspectives. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of the TITUS pipeline. TUs instances are segmented, then the ISO6436 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [1]

    Wagon and container codes detection and recognition based on yolov8

    Alejandro Diaz-Diaz, Franciso Parrilla, R De La Iglesia, Rafael Barea, and Luis M Bergasa. Wagon and container codes detection and recognition based on yolov8. In 2024 7th Iberian Robotics Conference (ROBOT), pages 1–6. IEEE, 2024

    work page 2024

  2. [2]

    SVTR: Scene text recognition with a single visual model

    Y Du, Z Chen, C Jia, X Yin, T Zheng, C Li, Y Du, and YG Jiang. SVTR: Scene text recognition with a single visual model. arxiv 2022.arXiv preprint arXiv:2205.00159, 2022

    arXiv 2022

  3. [3]

    Glenn Jocher et. al. ultralytics/yolov5: v6.0 - YOLOv5n ’Nano’ models, Roboflow integration, TensorFlow export, OpenCV DNN support, October 2021. URLhttps: //doi.org/10.5281/zenodo.5563715

    doi:10.5281/zenodo.5563715 2021

  4. [4]

    Freight containers — Coding, identification and marking

    Int’l Organization for Standardization. Freight containers — Coding, identification and marking. Standard, Geneva, CH, 2022

    work page 2022

  5. [5]

    From perception-action loops to imitation processes: A bottom-up approach of learning by imitation.Applied Artificial Intelligence, 12(7-8):701–727, 1998

    Philippe Gaussier, Sorin Moga, Mathias Quoy, and Jean-Paul Banquet. From perception-action loops to imitation processes: A bottom-up approach of learning by imitation.Applied Artificial Intelligence, 12(7-8):701–727, 1998

    work page 1998

  6. [6]

    Synthetic data for text local- isation in natural images

    Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text local- isation in natural images. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2315–2324, 2016

    work page 2016

  7. [7]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. GÜLSOYLU ET AL.: TRUDI AND TITUS11

    work page 2016

  8. [8]

    Mask R-CNN

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In Proc. of the IEEE international conference on computer vision, pages 2961–2969, 2017

    work page 2017

Show all 43 references
  1. [9]

    Build an effective system for container code recognition

    Nong Thi Hoa and Nguyen Van Tao. Build an effective system for container code recognition. InInternational Conference on Advances in Information and Communi- cation Technology, pages 32–39. Springer, 2023

    2023

  2. [10]

    Automatic recognition of container serial code

    Chung-Chian Hsu, Yu-Zen Yang, Arthur Chang, SM Salahuddin Morsalin, Guan-Ting Shen, and Li-Shin Shiu. Automatic recognition of container serial code. In2023 In- ternational Conference on Consumer Electronics-Taiwan (ICCE-Taiwan), pages 257–

  3. [11]

    ICDAR 2015 competition on robust reading

    Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, An- drew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. ICDAR 2015 competition on robust reading. In2015 13th international conference on docum...

    2015

  4. [12]

    MobileNetV3

    Brett Koonce. MobileNetV3. InConvolutional neural networks with swift for tensor- flow: image recognition and dataset categorization, pages 125–144. Springer, 2021

    2021

  5. [13]

    Automatic character recognition for moving and stationary vehicles and containers in real-life images

    John CM Lee. Automatic character recognition for moving and stationary vehicles and containers in real-life images. InIJCNN’99. International Joint Conference on Neural Networks. Proceedings (Cat. No. 99CH36339), volume 4, pages 2824–2828. IEEE, 1999

    1999

  6. [14]

    Towards end-to-end container code recogni- tion.Multimedia Tools and Applications, 81(11):15901–15918, 2022

    Yanchao Li, Hao Li, and Guangwei Gao. Towards end-to-end container code recogni- tion.Multimedia Tools and Applications, 81(11):15901–15918, 2022

    2022

  7. [15]

    Real-time scene text detection with differentiable binarization

    Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. InProc. of the AAAI Conf. on Artificial Intelligence, volume 34, pages 11474–11481, 2020

    2020

  8. [16]

    Real-time scene text detection with differentiable binarization and adaptive scale fusion.IEEE transac- tions on pattern analysis and machine intelligence, 45(1):919–931, 2022

    Minghui Liao, Zhisheng Zou, Zhaoyi Wan, Cong Yao, and Xiang Bai. Real-time scene text detection with differentiable binarization and adaptive scale fusion.IEEE transac- tions on pattern analysis and machine intelligence, 45(1):919–931, 2022

    2022

  9. [17]

    Review of scene text detection and recogni- tion.Archives of computational methods in engineering, 27(2):433–454, 2020

    Han Lin, Peng Yang, and Fanlong Zhang. Review of scene text detection and recogni- tion.Archives of computational methods in engineering, 27(2):433–454, 2020

    2020

  10. [18]

    Microsoft COCO: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ra- manan, Piotr Dollár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. InComputer vision–ECCV 2014: 13th European conference, zurich, Switzer- land, September 6-12, 2014, proceedi...

    2014

  11. [19]

    Inconsistent defect labels: Essence, causes, and influence

    Shiran Liu, Zhaoqiang Guo, Yanhui Li, Chuanqi Wang, Lin Chen, Zhongbin Sun, Yum- ing Zhou, and Baowen Xu. Inconsistent defect labels: Essence, causes, and influence. IEEE Transactions on Software Engineering, 49(2):586–610, 2022

    2022

  12. [20]

    Path aggregation network for instance segmentation

    Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 8759–8768, 2018. 12GÜLSOYLU ET AL.: TRUDI AND TITUS

    2018

  13. [21]

    Lightweight container number recognition based on deep learning.International Journal of System Assurance Engineering and Man- agement, pages 1–14, 2025

    Tao Liu, Xianqing Wu, and Fang Li. Lightweight container number recognition based on deep learning.International Journal of System Assurance Engineering and Man- agement, pages 1–14, 2025

    2025

  14. [22]

    ABCNet: Real-time scene text spotting with adaptive bezier-curve network

    Yuliang Liu, Hao Chen, Chunhua Shen, Tong He, Lianwen Jin, and Liangwei Wang. ABCNet: Real-time scene text spotting with adaptive bezier-curve network. Inpro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9809–9818, 2020

    2020

  15. [23]

    Swin transformer: Hierarchical vision transformer using shifted win- dows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted win- dows. InProceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  16. [24]

    Neural network application to container num- ber recognition

    Ho C Lui, Chung M Lee, and Fang Gao. Neural network application to container num- ber recognition. InProceedings., Fourteenth Annual International Computer Software and Applications Conference, pages 190–195. IEEE, 1990

    1990

  17. [25]

    RTMDet: An empirical study of designing real-time object detectors.arXiv preprint arXiv:2212.07784, 2022

    Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. RTMDet: An empirical study of designing real-time object detectors.arXiv preprint arXiv:2212.07784, 2022

    2022 arXiv

  18. [26]

    The impact of automatic identification on supply chain operations.The International Journal of Logistics Management, 2003

    Duncan McFarlane and Yossi Sheffi. The impact of automatic identification on supply chain operations.The International Journal of Logistics Management, 2003

    2003

  19. [27]

    Hoang-Sy Nguyen, Cong-Danh Huynh, and Nhat-Quan Bui. Digital transformation for shipping container terminals using automated container code recognition.TELKOM- NIKA (Telecommunication Computing Electronics and Control), 21(3):535–544, 2023

    2023

  20. [28]

    Contrast enhancement and brightness preservation using multi-decomposition histogram equalization.arXiv preprint arXiv:1307.3054, 2013

    Sayali Nimkar, Sanal Varghese, and Sucheta Shrivastava. Contrast enhancement and brightness preservation using multi-decomposition histogram equalization.arXiv preprint arXiv:1307.3054, 2013

    2013 arXiv

  21. [29]

    Robust container code recognition sys- tem

    Wei Pan, Yangsheng Wang, and Hongji Yang. Robust container code recognition sys- tem. InFifth World Congress on Intelligent Control and Automation (IEEE Cat. No. 04EX788), volume 5, pages 4061–4065. IEEE, 2004

    2004

  22. [30]

    SAR automatic target recognition based on multiview deep learning framework.IEEE Transactions on Geoscience and Remote Sensing, 56(4):2196–2210, 2017

    Jifang Pei, Yulin Huang, Weibo Huo, Yin Zhang, Jianyu Yang, and Tat-Soon Yeo. SAR automatic target recognition based on multiview deep learning framework.IEEE Transactions on Geoscience and Remote Sensing, 56(4):2196–2210, 2017

    2017

  23. [31]

    Optical container code recognition and its impact on the maritime supply chain

    Ravindra Shetty, Rebeca Cáceres, John Pastrana, and Luis Rabelo. Optical container code recognition and its impact on the maritime supply chain. InProceedings of the 2012 Industrial and Systems Engineering Research Conference, pages 1535–1544, 2012

    2012

  24. [32]

    RFID technology and its application to port-based container logistics.JOCEC, 21(4):332–347, 2011

    Xiaoning Shi, Dongkai Tao, and Stefan V oß. RFID technology and its application to port-based container logistics.JOCEC, 21(4):332–347, 2011

    2011

  25. [33]

    An overview of the Tesseract OCR engine

    Ray Smith. An overview of the Tesseract OCR engine. InNinth international confer- ence on document analysis and recognition (ICDAR 2007), volume 2, pages 629–633. IEEE, 2007. GÜLSOYLU ET AL.: TRUDI AND TITUS13

    2007

  26. [34]

    Drone-based identification of containers and semi-trailers in inland ports

    Jana Teegen, André Kelm, Ole Grasse, Maris Hillemann, Emre Gülsoylu, and Simone Frintrop. Drone-based identification of containers and semi-trailers in inland ports. EasyChair Preprint 14025, EasyChair, 2024

    2024

  27. [35]

    MobileOne: An improved one millisecond mobile backbone

    Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ran- jan. MobileOne: An improved one millisecond mobile backbone. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7907–7917, 2023

    2023

  28. [36]

    Lightweight container code recognition based on multi- reuse feature fusion and multi-branch structure merger.Journal of Real-Time Image Processing, 20(6):108, 2023

    Dapeng Yang, Guanghui Wang, Mingtang Liu, Shuang Yue, Hao Zhang, Xiaokang Chen, and Mengxiao Zhang. Lightweight container code recognition based on multi- reuse feature fusion and multi-branch structure merger.Journal of Real-Time Image Processing, 20(6):108, 2023

    2023

  29. [37]

    Advanced text detection of container numbers via dual-branch adaptive multi-scale network.Applied Sciences, 15(3):1492, 2025

    Li Yao, Chenchen Tang, and Yan Wan. Advanced text detection of container numbers via dual-branch adaptive multi-scale network.Applied Sciences, 15(3):1492, 2025

    2025

  30. [38]

    A two-stage auto- matic container code recognition method considering environmental interference.Ap- plied Sciences, 14(11):4779, 2024

    Meng Yu, Shanglei Zhu, Bao Lu, Qiang Chen, and Tengfei Wang. A two-stage auto- matic container code recognition method considering environmental interference.Ap- plied Sciences, 14(11):4779, 2024

    2024

  31. [39]

    Ro- bustScanner: Dynamically enhancing positional clues for robust text recognition

    Xiaoyu Yue, Zhanghui Kuang, Chenhao Lin, Hongbin Sun, and Wayne Zhang. Ro- bustScanner: Dynamically enhancing positional clues for robust text recognition. In European conference on computer vision, pages 135–151. Springer, 2020

    2020

  32. [40]

    A vertical text spotting model for trailer and container codes.IEEE Transactions on Instrumentation and Measurement, 70: 1–13, 2021

    Ran Zhang, Zhila Bahrami, and Zheng Liu. A vertical text spotting model for trailer and container codes.IEEE Transactions on Instrumentation and Measurement, 70: 1–13, 2021

    2021

  33. [41]

    A practical unified network for localization and recognition of arbitrary-oriented container code and type

    Jian Zhao, Ning Jia, Xianhui Liu, Gang Wang, and Weidong Zhao. A practical unified network for localization and recognition of arbitrary-oriented container code and type. IEEE Trans. on Instrumentation and Measurement, 2024

    2024

  34. [42]

    EAST: an efficient and accurate scene text detector

    Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. EAST: an efficient and accurate scene text detector. InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551–5560, 2017

    2017

  35. [43]

    De- formable DETR: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. De- formable DETR: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.