Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Velocity Completion Task and Method for Event-based Player Positional Data in Soccer

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A graph recurrent network reconstructs soccer player velocities from event-time positions, cutting the error by 58% versus the rule-based baseline.

desk verdict A genuinely new task with an honest evaluation; the main weakness is the underspecified rule-based baseline and missing error bars, both fixable. read the letter →

arxiv 2505.16199 v1 pith:6GSA55GA submitted 2025-05-22 cs.AI

Rikuhei Umemoto , Keisuke Fujii This is my paper
classification cs.AI
keywords velocitycompletioneventdatasocceranalyticsgraphrecurrentneuralnetworkdeeplearninginformationmissingspaceevaluation
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 defines a new task: estimating the velocities of all soccer players from event-based positional data, where only the coordinates of the ball and every player at discrete event times (passes, shots, and similar) are known. It argues that this is feasible with a graph recurrent neural network that models both the graph structure of player-to-player and player-to-ball interactions and the temporal dependencies across consecutive events. The authors report that the GRNN achieves a root mean square error of 1.90 m/s on the test events, a 58% reduction over the 4.51 m/s rule-based baseline, and that two space-evaluation metrics, the Potential Pitch Control Field (a per-location ball-control probability) and Off-Ball Scoring Opportunities, computed from the completed velocities, are closer to tracking-data ground truth than those from the baseline. If this holds, velocity-dependent tactical analysis could be performed on cheap, publicly available event data rather than expensive tracking data.

What carries the argument

The central object is the Graph Recurrent Neural Network (GRNN), which combines a graph neural network over the players and ball with a recurrent update over event sequences. At each event, node features are positions and flags (ball, player, attacker direction, ball holder, goalkeeper); edges encode teammate relations and ball-holder relations; a node-to-edge function produces edge features and an edge-to-node function updates node features. The resulting graph encoder produces latent states that feed a GRU-style recurrent hidden state across the k preceding events, and a graph decoder outputs velocities for all players. The design is the mechanism that lets the model use both spatial interaction structure and temporal dependencies simultaneously, which the experiments show is the combination that minimizes velocity error.

What would settle it

Compute the rule-based velocity baseline using the exact formula from the earlier rule-based work on the same 55-game dataset and test split; if its RMSE is at or below 1.90 m/s, the claimed improvement does not stand.

Watch

Extended reading notes

Core claim

The central claim is that player velocities, which are absent from event-based soccer data, can be completed directly from the positions of all players and the ball at event times, and that a deep learning model doing this outperforms the rule-based velocity-setting used in prior work. On the test events, the GRNN—a graph encoder-decoder coupled with a GRU-style recurrent update—reaches an RMSE of 1.90 m/s, while the rule-based baseline reaches 4.51 m/s; adding event-action type and end-of-event ball coordinates does not improve the GRNN further. When the completed velocities are plugged into the Potential Pitch Control Field (PPCF, the per-location probability of ball control) and Off-Ball Scoring Opportunities (OBSO, a measure of off-ball scoring threat) metrics, the resulting values are closer to those computed from ground-truth tracking data than the baseline's values, in aggregate and in the large majority of individual events. This establishes velocity completion as a viable intermediate step for event-data-driven team analysis.

Load-bearing premise

The rule-based velocity baseline used as the comparison target is treated as the state of the art, but its formula and parameters are not given here, so the reported 58% error reduction depends on that baseline being reasonably strong.

Editorial extensions

If this is right

  • Velocity-dependent analysis becomes applicable to event data for attacking-third events, without requiring full tracking data.
  • Space evaluation metrics such as PPCF and OBSO can be computed from completed velocities and yield values similar to tracking-data ground truth, enabling cheaper large-scale evaluation.
  • Information completion in sports is extended from filling missing positions to completing features entirely absent from the test dataset.
  • The combination of graph structure and temporal recurrence is what drives the improvement; the paper's ablations show GNN-only and RNN-only variants are less accurate.

Reading between the lines

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

  • The same architecture could be extended to all events on the pitch, not just the attacking third, which the authors list as future work; if it transfers, whole-match velocity-based metrics could be computed from event data alone.
  • Public event datasets that lack full player positions (for example some open data) would first need a position-completion stage; combining that with this velocity completion could yield a low-cost proxy for tracking data.
  • A simpler sanity-check baseline, such as finite-difference velocities from event positions with smoothing, is not tested in the paper; if that matched the GRNN, the need for deep learning would be weaker.
  • The method may transfer to other team sports with event data, where velocity-dependent space metrics already exist.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a new task: estimating the instantaneous velocities of all soccer players from event-time positional data alone (ball and player positions at event occurrences), where velocities are absent from the test data. The authors introduce a Graph Recurrent Neural Network (GRNN) that combines a GNN encoder/decoder with a GRU to exploit both graph structure and temporal dependencies. They compare against a rule-based baseline from their own prior work, as well as MLP, GNN, VAE, VGNN, RNN, VRNN, GVRNN, and variants with additional features (event action type and ball end position). Using 55 games of Japanese J1 League data, the GRNN achieves the lowest RMSE (1.90 m/s) against tracking-derived ground truth, outperforming the rule-based baseline (4.51 m/s). They also compute Potential Pitch Control Field (PPCF) and Off-Ball Scoring Opportunities (OBSO) errors and find that GRNN-complemented velocities yield values closer to ground truth than the rule-based baseline.

Significance. If the claims hold, the task is a practical contribution for analysts with only event data, as it enables velocity-dependent metrics like PPCF/OBSO without requiring full tracking data. The paper is honest about the architecture not being novel; the novelty lies in the task definition and the input/output modules. The experimental setup is partly reproducible: hyperparameters are listed in Appendix B, and a GitHub link to a reference implementation is given in Appendix A. However, the significance is currently tempered by three issues: the rule-based baseline is not specified, there is no uncertainty quantification on the model comparisons, and the loss function for the VAE-structured models contains a sign error that invalidates those baselines. With these fixed, the paper could provide a solid benchmark for velocity completion from event data.

major comments (4)
  1. [3.1 and 4.2.1] The rule-based baseline from Umemoto & Fujii (2023) is never specified; no formula, parameter values, or implementation details are provided. Since the paper's central quantitative claim is the 58% error reduction (Section 4.2.2, Table 2) relative to this baseline, the baseline must be described precisely or shown to be a non-trivial competitor. Without this, the comparisons in Tables 2-5 cannot be properly evaluated.
  2. [4.2.2, Tables 2-3] The paper reports a single averaged RMSE per model without error bars or significance tests. The statement "We obtained the average RMSE from 10 inferences in the test dataset" does not clarify whether these are 10 independent training runs or 10 stochastic forward passes. The difference between GRNN (1.90 m/s) and RNN (1.98 m/s) is small and could easily arise from seed-to-seed variation, so the claim that the proposed GRNN is superior to the RNN is not statistically supported. Tables 4 and 5 similarly report PPCF/OBSO errors without any measure of uncertainty across test events or model runs.
  3. [3.1, Equations (9)-(11)] The paper states "we do not have information on player jersey numbers to distinguish the players," yet the GRNN and all RNN-structured models process a sequence of events with a fixed set of player indices. If the ordering of players in the input x_t is not consistent across events, the recurrent hidden state h_t (Eq. 11) does not track the same physical player over time, undermining the temporal modeling that motivates the GRNN. The paper does not explain how player ordering is established in the data, nor how ground-truth velocities v_i_j are matched to predicted velocities for RMSE computation. This is load-bearing for the claimed advantage of the graph-plus-recurrent architecture over a pure GNN.
  4. [4.2.1, Equations (13) and (15)] For the VAE-structured models, the loss is defined as MSE - KL, where KL is the D_KL term. Minimizing this objective encourages the KL divergence to grow without bound, which is the opposite of the standard variational objective (reconstruction error + KL). This sign error means the VAE, VGNN, VRNN, and GVRNN implementations are not valid variational models, and their poor performance (Tables 2 and 3) cannot be attributed to the Gaussian assumption as the paper does in Appendix D. The loss should be corrected to MSE + KL (or the appropriate ELBO form) and the experiments rerun.
minor comments (5)
  1. [3.1] The sentence "player j shall only be judged if he is a teammate of another player, a goalkeeper, or the player in possession of the ball" is grammatically unclear and should be rephrased to explain which players are included in the task.
  2. [Appendix A] There is a typo in the GitHub link: "model modelhttps://github.com/keisuke198619/C-OBSO" should be a single proper URL.
  3. [4.2.1, VAE definition] The notation for the VAE distributions is inconsistent: the approximate posterior is written as q_phi(z_i | x_i, v_i, z_i), with z_i appearing on both sides; this likely should be q_phi(z_i | x_i, v_i). Please correct the notation.
  4. [Table 1] It would be helpful to explain why D* has fewer training and validation samples than D despite the same test sample count, since the paper only states that the series length increases the possibility of missing information.
  5. [4.2.2] The phrase "average RMSE from 10 inferences" should clarify whether these are 10 independent training runs with different random seeds or 10 stochastic forward passes of a single model.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the velocity completion model is evaluated on held-out test events against tracking-data ground truth, and the only self-citation is an underspecified rule-based baseline used as a comparison target, which does not force the reported result.

full rationale

The central derivation is non-circular. The model input is event-time positions, and the output velocities are compared with ground-truth velocities obtained from 25 Hz tracking data at event frames, with train/validation/test splits by games (Section 4.1, Tables 1-2). Thus the GRNN's 1.90 m/s RMSE is a genuine prediction on held-out test events, not a fitted value renamed as a prediction. The PPCF/OBSO comparison (Section 4.3) also uses ground-truth tracking velocities to define errors; no equation reduces the evaluation to the model's training objective. The only self-citation is the rule-based baseline from Umemoto and Fujii (2023), adopted in Sections 3.1 and 4.2.1 with no formula or parameters given in this paper. This is a transparency and benchmark-quality concern, but it is not circularity: the baseline is a comparison target, not an input from which the result is derived, and the core RMSE/PPCF/OBSO numbers are independently measured against tracking-data ground truth. The architecture is adapted from external work (Everett et al. 2023; Yeh et al. 2019), and the GVRNN reference implementation is linked externally. No step in the derivation chain is equivalent to its inputs by construction; the minor self-citation does not carry the central claim.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The claims rest on the model architecture borrowed from prior work, the proprietary dataset, the unverified ground-truth velocities, the rule-based baseline from the authors' prior paper, and the application metrics from Spearman (2018). No new physical entities are introduced; the main ledger items are hand-chosen hyperparameters and domain assumptions about data and baselines.

free parameters (5)
  • learning rate = 1e-4
    Fixed in Appendix B; chosen by hand rather than tuned on the validation set, and it affects convergence and final velocity error.
  • latent dimension num_hid = 8
    Fixed in Appendix B; model capacity choice that influences all MLP/GNN/VAE modules.
  • RNN hidden dimension rnn_dim = 46
    Fixed in Appendix B; affects the temporal modeling capacity of the RNN and GRNN models.
  • event sequence window k = 10
    Set in Section 3.1; the number of preceding events used by RNN-structured models, chosen by hand and affecting the temporal context available for velocity inference.
  • PPCF/OBSO model parameters (control rate, temporal uncertainty, reaction time, max speed) = GitHub defaults from LaurieOnTracking
    Appendix C imports these from prior work without fitting to this dataset; the space-evaluation conclusions in Section 4.3 depend on them.
assumptions (4)
  • domain assumption Ground-truth velocities from 25 Hz tracking data are accurate and correctly aligned with event times.
    Section 4.1 says velocities are calculated from tracking data at event frames, but no differentiation method, smoothing, or error analysis is given; all RMSE and space-evaluation benchmarks depend on this.
  • domain assumption The rule-based velocity baseline from Umemoto and Fujii (2023) is correctly implemented and representative of prior practice.
    Sections 3.1 and 4.2.1 adopt it as the comparison target without specifying its formula or parameters.
  • domain assumption The PPCF and OBSO models of Spearman (2018) are valid space evaluation metrics with appropriate default parameters.
    Section 4.3.1 uses them as the application benchmark, with parameters set to shared GitHub defaults from LaurieOnTracking (Appendix C).
  • domain assumption The train/validation/test split by consecutive games (40/8/7) means the held-out games are representative of the same distribution.
    Section 4.1 describes the split; potential season-long drift in tactics or data collection is not addressed, so the test-set results may not generalize to other leagues or seasons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Velocity Completion Task and Method for Event-based Player Positional Data in Soccer." pith.science (2026). https://pith.science/paper/6GSA55GA

@misc{pith2026250516199,
  author       = {Pith},
  title        = {Pith review of: Velocity Completion Task and Method for Event-based Player Positional Data in Soccer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6GSA55GA}},
  note         = {Machine review of arXiv:2505.16199}
}
read the original abstract

In many real-world complex systems, the behavior can be observed as a collection of discrete events generated by multiple interacting agents. Analyzing the dynamics of these multi-agent systems, especially team sports, often relies on understanding the movement and interactions of individual agents. However, while providing valuable snapshots, event-based positional data typically lacks the continuous temporal information needed to directly calculate crucial properties such as velocity. This absence severely limits the depth of dynamic analysis, preventing a comprehensive understanding of individual agent behaviors and emergent team strategies. To address this challenge, we propose a new method to simultaneously complete the velocity of all agents using only the event-based positional data from team sports. Based on this completed velocity information, we investigate the applicability of existing team sports analysis and evaluation methods. Experiments using soccer event data demonstrate that neural network-based approaches outperformed rule-based methods regarding velocity completion error, considering the underlying temporal dependencies and graph structure of player-to-player or player-to-ball interaction. Moreover, the space evaluation results obtained using the completed velocity are closer to those derived from complete tracking data, highlighting our method's potential for enhanced team sports system analysis.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

    Brefeld, U., Lasek, J., Mair, S. (2019). Probabilistic movement models and zones of control. Machine Learning, 108(1), 127–147,

    work page 2019

  2. [2]

    Caetano, F.G., Barbon Junior, S., Torres, R.d.S., Cunha, S.A., Ruffino, P.R.C., Mar- tins, L.E.B., Moura, F.A. (2021). Football player dominant region determined by a novel model based on instantaneous kinematics variables.Scientific reports, 11(1), 18209,

    work page 2021

  3. [3]

    Cao, W., Wang, D., Li, J., Zhou, H., Li, L., Li, Y. (2018). Brits: Bidirectional recurrent imputation for time series.Advances in neural information processing systems, 31, ,

    work page 2018

  4. [4]

    Capellera, G., Ferraz, L., Rubio, A., Agudo, A., Moreno-Noguer, F. (2024). Trans- portmer: A holistic approach to trajectory understanding in multi-agent sports. Proceedings of the asian conference on computer vision(pp. 1652–1670)

    work page 2024

  5. [5]

    Che, Z., Purushotham, S., Cho, K., Sontag, D., Liu, Y. (2018). Recurrent neural networks for multivariate time series with missing values. Scientific reports, 8(1), 6085,

    work page 2018

  6. [6]

    Cho, K. (2014). Learning phrase representations using rnn encoder-decoder for statistical machine translation.arXiv preprint arXiv:1406.1078, ,

    arXiv 2014

  7. [7]

    Choi, H., Kim, H., Lee, M., Kim, C.-J., Yoon, J., Ko, S.-K. (2024). Dbhp: Trajectory imputation in multi-agent sports using derivative-based hybrid prediction.arXiv preprint arXiv:2408.10878, ,

    work page Pith review arXiv 2024

  8. [8]

    Chung, J., Kastner, K., Dinh, L., Goel, K., Courville, A.C., Bengio, Y. (2015). A recur- rent latent variable model for sequential data.Advances in neural information processing systems, 28, ,

    work page 2015

Show all 43 references
  1. [9]

    Clevert, D.-A., Unterthiner, T., Hochreiter, S. (2015). Fast and accurate deep network learning by exponential linear units (elus).arXiv preprint arXiv:1511.07289, ,

    2015 arXiv

  2. [10]

    Everett, G., Beal, R.J., Matthews, T., Early, J., Norman, T.J., Ramchurn, S.D. (2023). Inferring player location in sports matches: Multi-agent spatial imputa- tion from limited observations.Proceedings of the 2023 international conference 20 on autonomous agents and multiagen...

    2023

  3. [11]

    Fernandez, J., & Bornn, L. (2018). Wide open spaces: A statistical technique for measuring space creation in professional soccer.Sloan sports analytics conference (Vol. 2018)

    2018

  4. [12]

    Fujii, K. (2025). Machine learning in sports: Open approach for next play analytics. Springer Nature

    2025

  5. [13]

    Fujii, K., Tsutsui, K., Scott, A., Nakahara, H., Takeishi, N., Kawahara, Y. (2024). Adaptive action supervision in reinforcement learning from real-world multi- agent demonstrations. International Conference on Agents and Artificial Intelligence (ICAART 2024), 2, 27-39,

    2024

  6. [14]

    Fujimura, A., & Sugihara, K. (2005). Geometric analysis and quantitative evaluation of sport teamwork.Systems and Computers in Japan, 36(6), 49–58,

    2005

  7. [15]

    Glorot, X., & Bengio, Y. (2010). Understanding the difficulty of training deep feed- forward neural networks.Proceedings of the thirteenth international conference on artificial intelligence and statistics(pp. 249–256). Iwashita,S.,Scott,A.,Umemoto,R.,Ding,N.,Fujii,K. (2024). ...

    2010 arXiv

  8. [16]

    Ji, S., Luo, J., Yang, X. (2020). A comprehensive survey on deep music generation: Multi-level representations, algorithms, evaluations, and future directions.arXiv preprint arXiv:2011.06801, ,

    2020 arXiv

  9. [17]

    Kavanagh, R., Di Michele, R., Oliveira, R., McDaid, K., Rhodes, D., Morgans, R. (2024). The relationships between distances covered above generic and rela- tive speed thresholds by male soccer players in english premier league matches across two competitive seasons. the effect...

    2024

  10. [18]

    Khemani, B., Patil, S., Kotecha, K., Tanwar, S. (2024). A review of graph neural networks: concepts, architectures, techniques, challenges, datasets, applications, and future directions.Journal of Big Data, 11(1), 18,

    2024

  11. [19]

    Kidger, P., Morrill, J., Foster, J., Lyons, T. (2020). Neural controlled differential equations for irregular time series. Advances in neural information processing 21 systems, 33, 6696–6707,

    2020

  12. [20]

    Kingma, D.P. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, ,

    2013 arXiv

  13. [21]

    Kipf, T.N., & Welling, M. (2016). Variational graph auto-encoders.arXiv preprint arXiv:1611.07308, ,

    2016 arXiv

  14. [22]

    Kono, R., & Fujii, K. (2024). Mathematical models for off-ball scoring prediction in basketball.International workshop on machine learning and data mining for sports analytics(pp. 41–54)

    2024

  15. [23]

    Liu, Y., Yu, R., Zheng, S., Zhan, E., Yue, Y. (2019). Naomi: Non-autoregressive multiresolution sequence imputation.Advances in neural information processing systems, 32, ,

    2019

  16. [24]

    Martens, F., Dick, U., Brefeld, U. (2021). Space and control in soccer.Frontiers in Sports and Active Living, 3, 676179,

    2021

  17. [25]

    Morrone, G., Michelsanti, D., Tan, Z.-H., Jensen, J. (2021). Audio-visual speech inpainting with deep learning.Icassp 2021-2021 ieee international conference on acoustics, speech and signal processing (icassp)(pp. 6653–6657)

    2021

  18. [26]

    Narizuka, T., Yamazaki, Y., Takizawa, K. (2021). Space evaluation in football games via field weighting based on tracking data.Scientific reports, 11(1), 5509,

    2021

  19. [27]

    (2022, May 23)

    Tuyls, K. (2022, May 23). Multiagent off-screen behavior prediction in football. Scientific Reports, 12(1), 8638, https://doi.org/10.1038/s41598-022-12547-0 Retrieved from https://doi.org/10.1038/s41598-022-12547-0

    2022 doi

  20. [28]

    Penn, M.J., Donnelly, C.A., Bhatt, S. (2023). Continuous football player tracking from discrete broadcast data.arXiv preprint arXiv:2311.14642, ,

    2023 arXiv

  21. [29]

    Qi, M., Qin, J., Wu, Y., Yang, Y. (2020). Imitative non-autoregressive modeling for trajectory forecasting and imputation.Proceedings of the ieee/cvf conference on computer vision and pattern recognition(pp. 12736–12745). 22

    2020

  22. [30]

    Rahimian, P., da Silva Guerra Gomes, D.G., Berkovics, F., Toka, L. (2022). Let’s penetrate the defense: A machine learning model for prediction and valuation of penetrative passes.International workshop on machine learning and data mining for sports analytics(pp. 41–52)

    2022

  23. [31]

    Robberechts, P., Van Roy, M., Davis, J. (2023). un-xpass: Measuring soccer player’s creativity.Proceedings of the 29th acm sigkdd conference on knowledge discovery and data mining(pp. 4768–4777)

    2023

  24. [32]

    Shan, S., Li, Y., Oliva, J.B. (2023). Nrtsi: Non-recurrent time series imputation. Icassp 2023-2023 ieee international conference on acoustics, speech and signal processing (icassp)(pp. 1–5)

    2023

  25. [33]

    Spearman, W. (2018). Beyond expected goals.Proceedings of the 12th mit sloan sports analytics conference(pp. 1–17)

    2018

  26. [34]

    Springham, M., Williams, S., Waldron, M., Burgess, D., Newton, R.U. (2020). Large reductions in match play physical performance variables across a professional football season with control for situational and contextual variables.Frontiers in Sports and Active Living, 2, 570937,

    2020

  27. [35]

    Taki, T., Hasegawa, J., Fukumura, T. (1996). Development of motion analysis system for quantitative evaluation of teamwork in soccer games.Proceedings of 3rd ieee international conference on image processing(Vol. 3, pp. 815–818)

    1996

  28. [36]

    Tashiro, Y., Song, J., Song, Y., Ermon, S. (2021). Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in neural information processing systems, 34, 24804–24816,

    2021

  29. [37]

    Teranishi, M., Tsutsui, K., Takeda, K., Fujii, K. (2023). Evaluation of creating scor- ing opportunities for teammates in soccer via trajectory prediction. Machine learning and data mining for sports analytics: 9th international workshop, mlsa 2022, grenoble, france, september...

    2023

  30. [38]

    Umemoto, R., Tsutsui, K., Fujii, K. (2022). Location analysis of players in uefa euro 2020 and 2022 using generalized valuation of defense by estimating probabilities. arXiv preprint arXiv:2212.00021, ,

    2022 arXiv

  31. [39]

    Polo- sukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., ... Polo- sukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30, , 23

    2017

  32. [40]

    Williams, R.J., & Zipser, D. (1989). A learning algorithm for continually running fully recurrent neural networks.Neural computation, 1(2), 270–280,

    1989

  33. [41]

    Xu, Y., & Fu, Y. (2024). Deciphering movement: Unified trajectory generation model for multi-agent.arXiv preprint arXiv:2405.17680, ,

    2024 arXiv

  34. [42]

    Yeh, R.A., Schwing, A.G., Huang, J., Murphy, K. (2019). Diverse generation for multi- agent sports games. Proceedings of the ieee/cvf conference on computer vision and pattern recognition(pp. 4610–4619). Yeung,C.,&Fujii,K. (2024). Astrategicframeworkforoptimaldecisionsinfootba...

    2019

  35. [43]

    Zhang, X., Zhai, D., Li, T., Zhou, Y., Lin, Y. (2023). Image inpainting based on deep learning: A review.Information Fusion, 90, 74–94, 24

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.