Pith. sign in

REVIEW 4 major objections 5 minor 24 references

A Novel Frame Identification and Synchronization Technique for Smartphone Visible Light Communication Systems Based on Convolutional Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes that a lightweight convolutional neural network can identify and synchronize frames in smartphone screen-to-camera visible light communication, reporting 98.74% classification accuracy and an estimated 85% computation…

desk verdict A plausible CNN classification study whose synchronization claim is contradicted by its own timing table, fixable but currently overclaimed. read the letter →

arxiv 2506.23004 v1 pith:SJR2NIXA submitted 2025-06-28 cs.CV eess.IV

Vaigai Nayaki Yokar , Hoa Le-Minh , Xicong Li , Wai Lok Woo , Luis Nero Alves , Stanislav Zvanovec , Tran The Son , Zabih Ghassemlooy This is my paper · ORCID
classification cs.CVeess.IV
keywords ConvolutionalneuralnetworksScreen-to-cameravisiblelightcommunicationFramesynchronizationidentificationOpticalcameraSmartphoneImageclassificationSupervisedlearning
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 seeks to solve one of the hardest practical problems in screen-to-camera visible light communication: the receiver does not know when a displayed frame is payload data and when it is a synchronization marker. The proposed answer is to treat frame identification as an image-classification task. A deliberately small convolutional neural network is trained on QR and ASCII frame images, including cropped and rotated versions, to tell data frames from overhead sync frames; the paper reports 98.74% overall accuracy across three classification experiments. If this works in a live link, the receiver can recognize overhead frames and skip decoding their payload, which the paper estimates as an 85% reduction in computation time per sync frame. That would matter because frame synchronization is a known bottleneck in smartphone-to-smartphone optical links.

What carries the argument

The load-bearing mechanism is the overhead-frame classifier: a convolutional neural network that converts synchronization into a per-frame label. The architecture is intentionally small, with 32 and 16 filters in the two convolutional layers, each $3\times3$, followed by $2\times2$ max pooling and a single 128-neuron dense layer, so that it can run on a smartphone receiver. Its task is to separate the overhead QR sync frames from the large 4,000-character QR data frames, so the receiver knows when a packet starts and can skip expensive decoding of sync frames. The quantitative engine of the claimed gain is Eq. (3), $x(t) = (T - T_{cnn})/T$, which converts the assumed CNN processing time of 5 ms against a 33.33 ms data-frame processing time into the reported 85% system gain.

What would settle it

Measure the real inference time of the described CNN on the same smartphone model used in the experiments while the receiver is running; if $T_{cnn}$ exceeds the frame interval or 5 ms, the 85% gain in Eq. (3) would not hold. Then transmit a known interleaved sequence of data and overhead frames and count misclassified boundaries under motion and changing light; with a 98.74% reported accuracy, roughly 1.26% of frames are misclassified, so the experiment should check whether that residual error concentrates at frame boundaries and corrupts packet alignment.

Watch

Extended reading notes

Core claim

The central claim is that a small CNN with two convolutional layers using $3\times3$ kernels and ReLU, a $2\times2$ max-pooling layer, a flatten layer, a 128-neuron fully connected layer, and a logistic output can classify $100\times100$ pixel S2C frames into the right category reliably enough to drive synchronization. Trained on 4,000 custom images (two QR data-frame types, one ASCII frame type, and one QR overhead frame type) with crop and rotation augmentation, the model reaches 98.60% accuracy for payload classification, 99.60% for distinguishing data frames from ASCII interference, and 98.00% for spotting overhead frames, for an average of 98.74%. The paper argues this classification capability is what lets the receiver detect the sync frames that mark packet boundaries, and uses Eq. (3) to estimate the resulting system gain: with a 33.33 ms per-frame computation time and a 5 ms CNN time, skipping payload processing for sync frames saves about 85%.

Load-bearing premise

The load-bearing premise is that 98.74% classification accuracy on still images, together with the assumed 5 ms CNN processing time, transfers to a live screen-to-camera link so that the receiver actually finds packet boundaries and realizes the estimated 85% computation saving.

Editorial extensions

If this is right

  • A receiver using this CNN can mark the start of each packet by recognizing overhead frames, removing the need for a separate preamble-detection circuit in screen-to-camera links.
  • Because sync frames are classified before payload decoding, the system can skip the QR-decoding step for those frames, cutting per-frame computation from 33.33 ms to 5 ms under the paper's assumptions.
  • The reported accuracy of 98.74% on cropped and rotated inputs suggests the classifier tolerates the blur, cropping, and rotation that occur during user mobility.
  • In the paper's comparison table, the proposed model exceeds the classification accuracy of SVM, LeNet-5, GoogLeNet, and ImageNet while keeping mean processing time at 237 ms, second-fastest of the five.
  • Frame identification as image classification offers a software-only path to synchronization, avoiding extra hardware on the smartphone receiver.

Reading between the lines

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

  • Because the paper's experiments stop at image classification, an end-to-end link test that transmits known data with periodic overhead frames and measures packet error rate or throughput is the direct next step needed to confirm the 85% computation saving reaches the user.
  • The 85% gain depends on the assumed $T_{cnn}$; without measuring inference latency on the target phone, the saving could be smaller in practice, or the CNN could fall behind the camera frame rate.
  • A classifier of this kind could be retrained to mark lost or duplicated frames caused by the mismatch between screen refresh and camera frame rates, turning frame-skip recovery into another classification task.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a supervised CNN for frame identification in smartphone screen-to-camera (S2C) visible light communication. The authors build a custom dataset of QR and ASCII frame images, train a small CNN to distinguish data frames from overhead/synchronization frames, and report an overall accuracy of 98.74% across three experiments. They further claim an 85% system gain in synchronization, computed from Eq. (3) as x(t)=(T-T_cnn)/T, by skipping payload processing when overhead frames are detected. A comparison against SVM, LeNet-5, GoogLeNet, and ImageNet is presented in Table IV.

Significance. If the classification result were reproducible and if the synchronization claim were backed by end-to-end measurements, a lightweight CNN that reliably separates overhead frames from data frames would be a useful component for OCC receivers and a positive step toward deep-learning-based synchronization on smartphones. The authors have assembled a custom image set that includes cropping, rotation, and blur, which is a constructive effort. However, the paper does not provide the evidence needed for the main claimed contribution: there are no synchronization metrics, no latency or throughput measurements on the live link, no measurement of the CNN's running time on the target phone, and no dataset or code release. The classification result is plausible but unverifiable under standard reproducibility expectations, and the synchronization gain claim contains an internal contradiction with the paper's own timing table.

major comments (4)
  1. [IV-B, Eq. (3), Table IV] The claimed 85% synchronization gain is contradicted by the paper's own timing table. Table IV reports a mean time of 237 ms for the proposed method. Substituting T=33.33 ms and T_cnn=237 ms into Eq. (3) gives (33.33 - 237)/33.33 = -6.11, i.e., a negative gain. The text states that computation time is reduced 'from 33.33 ms to 5 ms,' but no definition or measurement of the 5 ms value is provided, and no measurement of T_cnn on the Google Pixel 6 Pro appears anywhere in the manuscript. This contradiction is load-bearing because the abstract and the conclusion claim improved synchronization and system performance on the basis of this arithmetic estimate. The authors must either reconcile Table IV with Eq. (3) or remove the 85% gain claim until it is supported by direct measurement.
  2. [IV-B] The synchronization benefit is asserted but never measured. The paper reports no packet error rate, frame-sync success rate, throughput, or end-to-end latency on the live S2C link. Classification accuracy on the custom image set is not a synchronization metric, and the assumption that 98.74% accuracy translates to reliable frame-boundary recovery is untested. An end-to-end experiment with explicit synchronization metrics is required before any claim of improved synchronization can be accepted.
  3. [III] The architecture description is internally inconsistent and prevents reproduction. A 100x100 input passed through two 3x3 no-padding convolutional layers and one 2x2 max-pooling layer produces 48x48 feature maps (100 -> 98 -> 96 -> 48), not the stated '24 x 24' feature maps. The text also says convolution is performed 'without striding' but later refers to a 'stride of 12 x 12' in the pooling layer. These contradictions make it impossible to determine the actual network structure and undermine the claimed lightweight comparison.
  4. [IV-A, Table IV] The baseline comparison is not reproducible. No training or testing protocol is given for the four baselines, including dataset splits, augmentation settings, training epochs, and hardware. No error bars or repeated runs are reported. Two references do not point to the named models: [22] is a LiDAR semantic segmentation network, not LeNet-5, and [24] is a general deep-learning review, not GoogLeNet. 'ImageNet' as a model name is also ambiguous because ImageNet is a dataset. These issues make the superiority claim in Table IV unverifiable.
minor comments (5)
  1. [Fig. 8] Figure 8's caption duplicates the caption of Figure 7 ('Training and testing accuracy of overhead detection'), but the figure is referenced in Section IV-B as showing frame rate; the caption should be corrected.
  2. [IV-A.1] The accuracies reported as 'less than 0.48%' and '0.444%' appear to be fractions expressed as percentages; as written they are inconsistent with the training curves, which start near 0.5 and 0.45, respectively.
  3. [Table III] The average performance is reported to four decimal places with no confidence intervals; given the small per-class dataset (1,000 images), the variability across runs should be reported.
  4. [IV-B heading] The heading 'SYCHRONIZATION' contains a typo and should read 'SYNCHRONIZATION'.
  5. [II-C and IV-B] The notation T is used both for the receiver computation time and as the sampling interval base in Eq. (2), which is confusing; a clear separation of symbols is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: classification accuracy is measured on a held-out split, and the synchronization gain is an unsupported estimate rather than a circular derivation.

full rationale

The paper's central empirical claim is the 98.74% classification accuracy reported in Section IV-A. This accuracy is obtained by training the CNN on 75% of a purpose-built dataset and testing on the remaining 25%, as stated in Section IV: 'We utilized 75% of the dataset for CNN training and validation, reserving the remaining 25% for testing.' The accuracy is thus an independent held-out measurement, not a quantity that is re-derived from the model's own training inputs. No fitted parameter is relabeled as a prediction, and no equation in the derivation is equivalent by construction to the measured accuracy. The synchronization gain in Section IV-B, Eq. (3), x(t) = (T - Tcnn)/T, is computed from an assumed CNN processing time of 5 ms, yielding 85% by simple arithmetic. This is an unsupported assumption rather than a circular step: the paper does not claim to have measured Tcnn, nor does it fit Tcnn from the classification results and then predict the gain from the same data. The inconsistency between the assumed 5 ms and the Table IV mean time of 237 ms for the proposed method is a serious correctness and validation gap, but it is not circular. The paper's self-citations, principally references [1]-[7], are contextual prior works on S2C techniques and do not supply a load-bearing theorem or ansatz that forces the paper's conclusions. No uniqueness theorem from the authors is invoked, and no external result is renamed as a new finding. Overall, the derivation chain is not circular, though the synchronization benefit remains unvalidated.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on an unreleased dataset, hand-picked CNN hyperparameters, and an assumed CNN inference time. There are no new physical entities. The most consequential assumption is that classification accuracy implies synchronization performance, which is never tested end-to-end.

free parameters (5)
  • CNN trainable weights and biases = not reported
    All convolutional and fully connected parameters are fit to the training set; no count, values, or final weights are given, so the model is not reproducible from the text.
  • CNN hyperparameters = 32 and 16 filters, 3x3 kernel, 128 FC units, 20 epochs, Adam
    Chosen by hand without ablation or sensitivity analysis; the central accuracy claim depends on this exact configuration.
  • T (frame computation time) = 33.33 ms
    Used in Eq. (3) to compute the 85% system gain, but the paper elsewhere lists the Rx camera frame rate as 60 fps, which corresponds to about 16.7 ms per frame; the source of 33.33 ms is unclear.
  • Tcnn (CNN processing time) = 5 ms
    Assumed in Eq. (3) to estimate the 85% system gain; no timing measurement on the Google Pixel 6 Pro is reported.
  • Train/test split and augmentation parameters = 75%/25% split; crop and rotation ranges not specified
    The augmented dataset is central to the claimed robustness, but the augmentation ranges, counts, and split strategy are not fully described.
assumptions (3)
  • domain assumption Periodic overhead frames allow the receiver to identify the start of a packet stream and correct time offsets.
    Section II-C states this without an end-to-end demonstration; the synchronization claim rests on it.
  • ad hoc to paper High CNN classification accuracy on the custom image dataset is sufficient for reliable frame synchronization in the live S2C link.
    No packet-error-rate, timing-offset, or throughput measurements connect the 98.74% classification accuracy to synchronization performance.
  • domain assumption The custom dataset represents real S2C channel conditions including blur, cropping, and rotation.
    The dataset is created from scratch with augmented images, but blur, mobility, and lighting conditions are not quantified, so representativeness is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Frame Identification and Synchronization Technique for Smartphone Visible Light Communication Systems Based on Convolutional Neural Networks." pith.science (2026). https://pith.science/paper/SJR2NIXA

@misc{pith2026250623004,
  author       = {Pith},
  title        = {Pith review of: A Novel Frame Identification and Synchronization Technique for Smartphone Visible Light Communication Systems Based on Convolutional Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJR2NIXA}},
  note         = {Machine review of arXiv:2506.23004}
}
read the original abstract

This paper proposes a novel, robust, and lightweight supervised Convolutional Neural Network (CNN)-based technique for frame identification and synchronization, designed to enhance short-link communication performance in a screen-to-camera (S2C) based visible light communication (VLC) system. Developed using Python and the TensorFlow Keras framework, the proposed CNN model was trained through three real-time experimental investigations conducted in Jupyter Notebook. These experiments incorporated a dataset created from scratch to address various real-time challenges in S2C communication, including blurring, cropping, and rotated images in mobility scenarios. Overhead frames were introduced for synchronization, which leads to enhanced system performance. The experimental results demonstrate that the proposed model achieves an overall accuracy of approximately 98.74%, highlighting its effectiveness in identifying and synchronizing frames in S2C VLC systems.

Figures

Figures reproduced from arXiv: 2506.23004 by the authors.

Figure 1
Figure 1. Smartphone-based VLC system: (a) Tx and Rx pair, and (b) functional system block diagram In OOK or color shift keying (CSK)-based systems, these cells are colored in white and black for '0's and '1's, respectively. A bit ‘1’ is simply represented as an optical pulse that occupies the entire part of the bit duration while a bit ‘0’ is represented by the absence of an optical pulse. For OOK modulation in S2C systems, … view at source ↗
Figure 2
Figure 2. Frame rate: (a) Tx (@120 fps), and (b) Rx (@60 fps). The asynchronous transmission in S2C communication leads to (i) latency considerations, (ii) frame skipping or duplications, and (iii) resilience to timing variations. To address this issue, in this work, overhead frames will be transmitted periodically, which helps the Rx to identify the beginning of a stream of packets as well as to correct any time offset. This… view at source ↗
Figure 3
Figure 3. S2C frame model The traditional method to separate the random data frames from the overhead frames can be computationally extensive. To reduce computational power and system complexity, we [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Proposed CNN model. Here, the CNN algorithm was created to classify 100 × 100 image size QR and ASCII codes. The image is convolved using three separate kernels, each of size 3 × 3, without striding in the convolution layer. A kernel size of "3 × 3" is a common choice …
Figure 5
Figure 5. Figure 5: Training and testing accuracy of payload detection. 2) Data frame detection The second experiment focuses on ensuring the system exclusively recognizes data frames (QR codes), filtering out other barcodes like ASCII codes. The CNN model is trained on various barcodes t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 14 canonical work pages

  1. [22]

    LENet: Lightweight And Efficient LiDAR Semantic Segmentation Using Multi-Scale Convolution Attention

    B. Ding, “LENet: Lightweight And Efficient LiDAR Semantic Segmentation Using Multi-Scale Convolution Attention,” arXiv preprint arXiv:2301.04275, 2023

    work page Pith review arXiv 2023

  2. [24]

    Deep learning,

    Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  3. [1]

    A Novel Blur Reduction Technique For QR And ASCII Coding In Smartphone Visible Light Communications,

    V. N. Yokar, Hoa-Le-Minh, F. Ghassemlooy, and W. L. Woo, “A Novel Blur Reduction Technique For QR And ASCII Coding In Smartphone Visible Light Communications,” in 2022 13th International Symposium on Communication Systems, Networks and Digital Signal Processing (CSNDSP), Jul. 2022, pp. 428–433. doi: 10.1109/CSNDSP54353.2022.9907993

    arXiv 2022

  4. [2]

    Performance evaluation technique for screen-to-camera-based optical camera communications,

    V. N. Yokar, H. Le-Minh, Z. Ghassemlooy, and W. L. Woo, “Performance evaluation technique for screen-to-camera-based optical camera communications,” IET Optoelectronics, vol. n/a, no. n/a, Aug. 2023, doi: 10.1049/ote2.12102

    doi:10.1049/ote2.12102 2023

  5. [3]

    Smartphone Beam Profile in a Screen-to- Camera-Based Optical Communication System,

    V. N. Yokar et al., “Smartphone Beam Profile in a Screen-to- Camera-Based Optical Communication System,” in 2023 17th International Conference on Telecommunications (ConTEL), Jul. 2023, pp. 1–6. doi: 10.1109/ConTEL58387.2023.10199032

    arXiv 2023

  6. [4]

    Fast Link Recovery via PTP-synchronized Nanosecond Optical Switching,

    V. Yokar et al., “Fast Link Recovery via PTP-synchronized Nanosecond Optical Switching,” arXiv preprint arXiv:2412.13778, 2024

    arXiv 2024

  7. [5]

    Non-Blind Image Restoration Technique in Screen–to–Camera based Optical Camera Communications,

    V. N. Yokar, H. Le-Minh, L. N. Alves, S. Zvanovec, W. L. Woo, and Z. Ghassemlooy, “Non-Blind Image Restoration Technique in Screen–to–Camera based Optical Camera Communications,” presented at the 2024 7th International Balkan Conference on Communications and Networking (BalkanCom), IEEE, 2024, pp. 101–106

    work page 2024

  8. [6]

    Data Detection Technique for Screen-to-Camera Based Optical Camera Communications,

    V. N. Yokar, H. Le-Minh, Z. Ghassemlooy, and W. L. Woo, “Data Detection Technique for Screen-to-Camera Based Optical Camera Communications,” presented at the 2024 14th International Symposium on Communication Systems, Networks and Digital Signal Processing (CSNDSP), IEEE, 2024, pp. 233–237

    work page 2024

Show all 24 references
  1. [7]

    Hemoglobin Estimation through Image Processing of Eye Image via Determination of the Density Parameter of a 2-D Wavelet Processed Image

    V. N. Yokar, “Hemoglobin Estimation through Image Processing of Eye Image via Determination of the Density Parameter of a 2-D Wavelet Processed Image”

  2. [8]

    Information Revealed by Vision: A Review on the Next-Generation OCC Standard for AIoV,

    N. -N. Dao, T. -H. Do, S. Cho, and S. Dustdar, “Information Revealed by Vision: A Review on the Next-Generation OCC Standard for AIoV,” IT Professional, vol. 24, no. 4, pp. 58–65, Aug. 2022, doi: 10.1109/MITP.2022.3180354

    2022

  3. [9]

    COBRA: Color barcode streaming for smartphone systems,

    T. Hao, R. Zhou, and G. Xing, “COBRA: Color barcode streaming for smartphone systems,” presented at the Proceedings of the 10th international conference on Mobile systems, applications, and services, 2012, pp. 85–98

    2012

  4. [10]

    VINCE: Exploiting visible light sensing for smartphone-based NFC systems,

    J. Niu, F. Gu, R. Zhou, G. Xing, and W. Xiang, “VINCE: Exploiting visible light sensing for smartphone-based NFC systems,” in 2015 IEEE Conference on Computer Communications (INFOCOM), May 2015, pp. 2722–2730. doi: 10.1109/INFOCOM.2015.7218664

    2015

  5. [11]

    NECAS: Near field communication system for smartphones based on visible light,

    J. Niu, W. Song, C. Liu, L. Shu, and C. Chen, “NECAS: Near field communication system for smartphones based on visible light,” in 2014 IEEE Wireless Communications and Networking Conference (WCNC), Apr. 2014, pp. 2426–2431. doi: 10.1109/WCNC.2014.6952729

    2014

  6. [12]

    Rain Bar: Robust Application-Driven Visual Communication Using Color Barcodes,

    Q. Wang, M. Zhou, K. Ren, T. Lei, J. Li, and Z. Wang, “Rain Bar: Robust Application-Driven Visual Communication Using Color Barcodes,” in 2015 IEEE 35th International Conference on Distributed Computing Systems, Jul. 2015, pp. 537–546. doi: 10.1109/ICDCS.2015.61

    2015 doi

  7. [13]

    SoftLight: Adaptive visible light communication over screen-camera links,

    W. Du, J. C. Liando, and M. Li, “SoftLight: Adaptive visible light communication over screen-camera links,” in IEEE INFOCOM 2016 - The 35th Annual IEEE International Conference on Computer Communications, Apr. 2016, pp. 1–9. doi: 10.1109/INFOCOM.2016.7524510

    2016

  8. [14]

    TETRIS: Smartphone-to-Smartphone Screen-Based Visible Light Communication,

    M. Stafford, A. Rogers, S. Wu, C. Carver, N. S. Artan, and Z. Dong, “TETRIS: Smartphone-to-Smartphone Screen-Based Visible Light Communication,” in 2017 IEEE 14th International Conference on Mobile Ad Hoc and Sensor Systems (MASS), Oct. 2017, pp. 570–574. doi: 10.1109/MASS.2017.101

    2017 doi

  9. [15]

    Smartphone to Smartphone Visible Light Communications,

    R. Boubezari, “Smartphone to Smartphone Visible Light Communications,” Northumbria University, 2018

    2018

  10. [16]

    Cognitive vision communication based on LED array and image sensor,

    J. -H. Yoo and S. -Y. Jung, “Cognitive vision communication based on LED array and image sensor,” in 2013 IEEE 56th International Midwest Symposium on Circuits and Systems (MWSCAS), Aug. 2013, pp. 1228–1231. doi: 10.1109/MWSCAS.2013.6674876

    2013

  11. [17]

    Lightsync: Unsynchronized visual communication over screen-camera links,

    W. Hu, H. Gu, and Q. Pu, “Lightsync: Unsynchronized visual communication over screen-camera links,” presented at the Proceedings of the 19th annual international conference on Mobile computing & networking, 2013, pp. 15–26

    2013

  12. [18]

    Visual light landmarks for mobile devices,

    N. Rajagopal, P. Lazik, and A. Rowe, “Visual light landmarks for mobile devices,” in IPSN-14 Proceedings of the 13th International Symposium on Information Processing in Sensor Networks, Apr. 2014, pp. 249–260. doi: 10.1109/IPSN.2014.6846757

    2014

  13. [19]

    Real-time screen-camera communication behind any scene,

    T. Li, C. An, X. Xiao, A. T. Campbell, and X. Zhou, “Real-time screen-camera communication behind any scene,” presented at the Proceedings of the 13th Annual International Conference on Mobile Systems, Applications, and Services, 2015, pp. 197–211

    2015

  14. [20]

    Gulli and S

    A. Gulli and S. Pal, Deep learning with Keras. Packt Publishing Ltd., 2017

    2017

  15. [21]

    Deep learning using linear support vector machines,

    Y. Tang, “Deep learning using linear support vector machines,” arXiv preprint arXiv:1306.0239, 2013

    2013 arXiv

  16. [23]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L. -J. Li, Kai Li, and Li Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition, Jun. 2009, pp. 248–255. doi: 10.1109/CVPR.2009.5206848

    2009

Pith tools

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