Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Genetic Algorithms for the Optimization of Diffusion Parameters in Content-Based Image Retrieval

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Genetic algorithms can tune diffusion-based image retrieval faster than grid search while reaching equal or better accuracy.

desk verdict A potentially useful GA-based tuner for CBIR diffusion parameters, but the diffusion equation as written makes α irrelevant to ranking, so the paper needs correction and code before its results can be trusted. read the letter →

arxiv 1908.06896 v1 pith:SIMP77SL submitted 2019-08-19 cs.CV

Federico Magliani , Laura Sani , Stefano Cagnoni , Andrea Prati This is my paper · ORCID
classification cs.CV
keywords geneticalgorithmsdiffusionprocesscontent-basedimageretrievalparametertuningkNNgraphmeanaverageprecisionparticleswarmoptimizationrandomwalk
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

Diffusion-based image retrieval improves accuracy by exploiting the manifold structure of a kNN graph, but its performance depends on several hand-set parameters. This paper claims those parameters can be tuned automatically with a genetic algorithm, using retrieval mean average precision (mAP) as the fitness function, and that this approach is faster than brute-force grid search, random search, and particle swarm optimization while matching or exceeding their accuracy. On Oxford5k, the genetic algorithm reached 94.44% mAP with 5,000 fitness evaluations in about 17,700 seconds, while grid search needed 200,000 evaluations and about 1,036,800 seconds for 94.43% mAP. If correct, the result removes a major practical obstacle to using diffusion retrieval on new datasets.

What carries the argument

The load-bearing object is the diffusion equation system $Ax=b$ on a kNN graph, where $A$ is the affinity matrix of database images, $b$ is the query vector, and $x$ is the ranking vector. The matrix and query are modified by the tunable operations $A = A - \alpha A$, $a_{ij} = a_{ij}^{\beta}$, and $b_i = b_i^{\gamma}$, while $k_s$, $k$, the iteration limit, and the truncation size bound the random walk and solver. A genetic algorithm is applied to these seven parameters, with mAP as the fitness function, so each fitness evaluation is a complete retrieval run. The mechanism's work is to turn parameter tuning into a directed evolutionary search, avoiding the exhaustive grid over all parameter combinations.

What would settle it

Run the same retrieval configuration twice with $\alpha$ set to two distinct values, for example 0.1 and 0.9, keeping all other tuned parameters fixed, and compare the ranked retrieval lists; if the rankings are identical, $\alpha$ has no effect on retrieval quality and the paper's tuned $\alpha$ values do not support its central claim about optimizing diffusion parameters.

Watch

Extended reading notes

Core claim

The paper's central claim is that a genetic algorithm can replace manual or exhaustive configuration of diffusion parameters in content-based image retrieval. The diffusion step is modelled as solving $Ax=b$ on an approximate kNN graph, where seven parameters control the walk and solver: the damping factor $\alpha$, the matrix exponent $\beta$, the query exponent $\gamma$, the number of random-walk steps $k_s$, the number of neighbours $k$, the solver iteration limit, and the truncation size. Treating the retrieval mAP as fitness, the genetic algorithm searches this space with tournament selection, single-point crossover, and mutation, keeping the best individuals found during the whole run. In the reported experiments, the genetic algorithm achieved 94.44% mAP on Oxford5k, 97.32% on Paris6k, and 94.20% on Oxford105k, beating particle swarm optimization and random search, and matching grid search with far fewer fitness evaluations. The paper also finds that the optimal diffusion parameters differ across datasets, so the tuning must be repeated per dataset.

Load-bearing premise

The load-bearing premise is that the diffusion process is correctly described by the equations given in the paper, and specifically that the damping parameter $\alpha$ genuinely changes the ranking produced by $Ax=b$; if $\alpha$ only scales $A$, the ranking is unchanged and all reported tuned $\alpha$ values would be meaningless.

Editorial extensions

If this is right

  • On Oxford5k, grid search reaches 94.43% mAP after 200,000 fitness evaluations, while the genetic algorithm reaches 94.44% after 5,000 evaluations, a 40-fold reduction in evaluations and a roughly 58-fold reduction in wall-clock time in the paper's experiments.
  • On Paris6k and Oxford105k, the genetic algorithm improves mAP over the manual literature configuration by 0.31 and 1.70 percentage points, respectively.
  • A user can adopt diffusion retrieval on a new dataset without hand-tuning, at the cost of one genetic-algorithm run.
  • Because the best diffusion parameters differ across datasets, the paper does not claim one universal configuration; the tuning must be repeated for each dataset.
  • The comparison framework means the genetic algorithm's advantage should grow when the parameter ranges widen, since evolutionary search samples a subset of the combinatorics rather than enumerating them.

Reading between the lines

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

  • If the equations are implemented exactly as written, $\alpha$ may be only a global scaling of $A$; the ranking recovered from $Ax=b$ would then be invariant to it, making the tuned $\alpha$ values a by-product of the search rather than a causal driver of accuracy. This can be checked by fixing the other six parameters and varying $\alpha$ alone.
  • The paper's fitness evaluation dominates runtime, so using a cheaper surrogate for mAP, such as a small subset of queries, could accelerate the same genetic algorithm; this extension is not explored in the paper.
  • Because the genetic algorithm's speed advantage comes from sampling the search space, testing it on larger datasets or with wider parameter bounds would make the gap over grid search explicit and test the scalability claim.
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 manuscript proposes using genetic algorithms (GAs) to automatically tune seven parameters of a diffusion-based image retrieval system built on approximate kNN graphs. The authors compare GA against grid search, random search, particle swarm optimization (PSO), and a manual configuration on Oxford5k, and against the manual configuration only on Paris6k and Oxford105k. On Oxford5k, they report that GA reaches 94.44% mAP after 5,000 fitness evaluations, slightly outperforming PSO (94.30%) and grid search (94.43%) while reducing wall-clock time, and substantially outperforming random search (93.67%) and manual configuration (90.95%). The paper also includes a sensitivity analysis of GA meta-parameters.

Significance. If the reported results are substantiated, the paper would offer a practical and relatively low-cost way to configure diffusion-based CBIR systems, which is a relevant problem since manual tuning is time-consuming and dataset-dependent. The authors' use of the external mAP as the fitness function avoids the circularity that would arise from optimizing a proxy objective, and the five-run repeatability check for the GA is a positive sign. However, the current significance is limited by the unresolved diffusion-equation ambiguity, the dataset-specific tuning of GA meta-parameters, and the incomplete comparisons on the other two datasets.

major comments (4)
  1. [Section 4.1] The equation A = A − αA implies A' = (1−α)A, so solving A'x = b yields x' = x/(1−α). Since all components of the ranking vector are multiplied by the same positive constant, the ranking and hence mAP are invariant to α. The paper nonetheless reports dataset-dependent optimal values of α (0.97, 0.87, 0.97) and likens α to the PageRank damping factor; in PageRank, α enters through a restart term such as x ← αWx + (1−α)b, not through a global rescaling of A. As written, the optimization of α is not well-defined. Please state the exact iterative update used (e.g., x_{t+1} = αAx_t + (1−α)b) or release the code; otherwise the reported tuned parameter values and the mAP comparisons are not reproducible from the text.
  2. [Section 5.2, Tables 1–6] The GA's meta-parameters (Gen, Pop, CxPb, MutPb, IndPb) were selected by maximizing mAP on Oxford5k, and the same Oxford5k test set was then used for the final comparison in Table 6. This gives GA an advantage over PSO and random search, whose meta-parameters were fixed a priori, and it likely inflates the reported 94.44% mAP. To make the comparison fair, the GA meta-parameters should be chosen on a separate validation set (or via a nested procedure), and the final evaluation should be reported on a held-out set.
  3. [Table 6] Only the GA is evaluated with multiple runs (avg = 94.39%, stdev = 0.038, max = 94.44%), while the PSO and random-search results are reported as single runs. Since PSO and random search are also stochastic, the observed differences (94.44% vs 94.30% vs 93.67%) may not be statistically significant. Please report repeated runs and dispersion for all stochastic baselines, or perform a significance test, before claiming that GA achieves the same or better results.
  4. [Sections 5.2–5.4, Tables 6–8] The claimed speed and accuracy advantage of GA over grid search, random search, and PSO is demonstrated only on Oxford5k. Tables 7 and 8 for Paris6k and Oxford105k compare only GA with the manual configuration, so the abstract's statement that a comparison with 'these methods' was made on three datasets is unsupported. Additionally, the wall-clock times in Table 6 imply very different per-evaluation costs (GA ≈ 3.5 s/eval, random search ≈ 1.35 s/eval, grid search ≈ 5.2 s/eval), which is unexplained and may confound the 'faster' claim; please report per-evaluation costs or use fixed wall-clock budgets.
minor comments (5)
  1. [Section 3.2] The phrase 'system of equation' should be 'system of equations', and the notation Ax = b should be introduced more explicitly with A, x, and b defined before the equation is used.
  2. [Section 4.1] The notation 'aij = aβij' is ambiguous; please write a_{ij} ← a_{ij}^β or use explicit superscripts to indicate element-wise exponentiation.
  3. [Section 5.1] The definition of Oxford105k is vague; specify the exact total number of database images (e.g., 5,063 + 100,000) and the query set used for evaluation.
  4. [Section 5.2, Tables 1–5] The statement 'Since these configurations lead to the same mAP (94.40%), the remaining parameters have been varied starting from the configuration which is fastest to compute' is not supported by the tables, which report only mAP and not runtime; include runtime or remove the 'fastest' justification.
  5. [Table 6] The 'manual configuration' row is presented as a baseline from the literature, but reference [27] is the authors' own prior work; clarify this and justify why it is a representative manual baseline for comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the GA's fitness is the external mAP, and reported results are optimization outcomes rather than fitted predictions.

full rationale

The paper's central claim is that a genetic algorithm can find diffusion parameter settings that match or beat grid, random, and PSO search on the same retrieval objective. This is not circular: the fitness function is the externally computed mean Average Precision of the diffusion retrieval pipeline, so the GA does not optimize a self-defined surrogate and then relabel it as the reported evaluation metric. The reported mAP for each method is simply the objective value of the best configuration found, which is the standard and fair meaning of a parameter-optimization comparison. The 'manual configuration' baseline taken from the authors' prior work [27] is a fixed comparison point, not a fitted parameter in this paper, and the approximate kNN graph construction from [27] is a reused component rather than a circular justification of the GA's efficiency claim. The tuning of GA meta-parameters (population size, crossover, mutation) on Oxford5k and then reporting the resulting Oxford5k mAP is an in-sample selection issue that can inflate the apparent result, but it is not a by-construction reduction: the reported mAP is not defined in terms of those meta-parameters. Separately, Section 4.1 contains a serious mathematical reproducibility problem because A = A − αA makes α a global scale factor, so the ranking solution x = (1−α)^{-1} A^{-1} b is independent of α; this is a correctness flaw about whether the written diffusion process matches the implementation, not a circular dependency between an input and a prediction. No quoted step exhibits a fitted parameter renamed as a prediction or an equation equal to its own input, so the circularity score is 0.

Assumptions & free parameters 8 free parameters · 3 assumptions · 0 invented entities

The central claim depends on the diffusion pipeline imported from prior work [21, 27, 40], on the chosen parameter ranges, and on the GA meta-parameters. All seven diffusion parameters are fitted to the test data; the paper does not provide an external validation set. No new entities are introduced.

free parameters (8)
  • alpha (α) = 0.97 (Oxford5k), 0.87 (Paris6k), 0.97 (Oxford105k)
    Damping factor in the diffusion system. In Section 4.1, A is replaced by A-alpha*A, which is a global scaling that may not affect the ranking solution, yet it is optimized per dataset.
  • beta (β) = 3, 1, 2
    Exponent applied to affinity matrix entries to remove weak neighbors. Fitted for each dataset.
  • gamma (γ) = 2, 2, 1
    Exponent applied to query vector entries. Fitted for each dataset.
  • ks = 53, 40, 68
    Number of random walk steps on the graph. Fitted per dataset.
  • k = 9, 11, 7
    Number of neighbors in the kNN graph. Fitted per dataset.
  • iterations = 10 (all datasets)
    Maximum iterations for the linear system solver. Fitted.
  • trunc = 4136, 3761, 18353
    Number of database elements used during diffusion. Fitted per dataset.
  • GA configuration (Gen, Pop, CxPb, MutPb, IndPb) = 100, 50, 0.3 or 0.5, 0.2, 0.1
    Meta-parameters of the genetic algorithm, chosen by experiments on Oxford5k. They affect search quality but are selected without a validation split.
assumptions (3)
  • domain assumption The diffusion process described in Sections 3 and 4, based on Zhou et al. [40] and Iscen et al. [21], correctly models manifold ranking for image retrieval.
    The paper adopts this method from prior work and does not verify its correctness.
  • domain assumption The approximate kNN graph construction from Magliani et al. [27] preserves retrieval quality after diffusion.
    The experiments rely entirely on this graph construction, which is cited but not described or verified in this paper.
  • ad hoc to paper The parameter ranges in Section 4.2 contain a near-optimal configuration for each dataset.
    The ranges are chosen by the authors and are not justified with sensitivity analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Genetic Algorithms for the Optimization of Diffusion Parameters in Content-Based Image Retrieval." pith.science (2026). https://pith.science/paper/SIMP77SL

@misc{pith2026190806896,
  author       = {Pith},
  title        = {Pith review of: Genetic Algorithms for the Optimization of Diffusion Parameters in Content-Based Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SIMP77SL}},
  note         = {Machine review of arXiv:1908.06896}
}
read the original abstract

Several computer vision and artificial intelligence projects are nowadays exploiting the manifold data distribution using, e.g., the diffusion process. This approach has produced dramatic improvements on the final performance thanks to the application of such algorithms to the kNN graph. Unfortunately, this recent technique needs a manual configuration of several parameters, thus it is not straightforward to find the best configuration for each dataset. Moreover, the brute-force approach is computationally very demanding when used to optimally set the parameters of the diffusion approach. We propose to use genetic algorithms to find the optimal setting of all the diffusion parameters with respect to retrieval performance for each different dataset. Our approach is faster than others used as references (brute-force, random-search and PSO). A comparison with these methods has been made on three public image datasets: Oxford5k, Paris6k and Oxford105k.

Figures

Figures reproduced from arXiv: 1908.06896 by the authors.

Figure 1
Figure 1. Two data distributions (red and orange dots). In this [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Enhancing Visual Re-ranking through Denoising Nearest Neighbor Graph via Continuous CRF

    cs.CV 2024-12 conditional novelty 6.0 of 10

    A clique-based C-CRF denoising of the nearest-neighbor affinity graph improves three graph-based visual re-ranking methods on ROxford and RParis.

Reference graph

Works this paper leans on

42 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Artem Babenko and Victor Lempitsky. 2015. Aggregating local deep features for image retrieval. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition . 1269–1277

    work page 2015

  2. [2]

    Thomas Bäck and Hans-Paul Schwefel. 1993. An overview of evolutionary algorithms for parameter optimization. Evolutionary computation 1, 1 (1993), 1–23

    work page 1993

  3. [3]

    James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research 13, Feb (2012), 281–305

    2012

  4. [4]

    James S Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. 2011. Al- gorithms for hyper-parameter optimization. In Advances in neural information processing systems. 2546–2554

    work page 2011

  5. [5]

    Wei Dong, Charikar Moses, and Kai Li. 2011. Efficient k-nearest neighbor graph construction for generic similarity measures. In Proceedings of the 20th Interna- tional Conference on World Wide Web. ACM, 577–586

    work page 2011

  6. [6]

    Ágoston E Eiben, Robert Hinterding, and Zbigniew Michalewicz. 1999. Parameter control in evolutionary algorithms.IEEE Transactions on evolutionary computation 3, 2 (1999), 124–141

    work page 1999

  7. [7]

    A. E. Eiben and James E. Smith. 2015. Introduction to Evolutionary Computing (2nd ed.). Springer Publishing Company, Incorporated

    work page 2015

  8. [8]

    Engelbrecht

    Andries P. Engelbrecht. 2007. Computational Intelligence: An Introduction (2nd ed.). Wiley Publishing

    work page 2007

Show all 42 references
  1. [9]

    Stefan Falkner, Aaron Klein, and Frank Hutter. 2018. Bohb: Robust and efficient hyperparameter optimization at scale. arXiv preprint arXiv:1807.01774 (2018)

    2018 arXiv

  2. [10]

    Félix-Antoine Fortin, François-Michel De Rainville, Marc-André Gardner, Marc Parizeau, and Christian Gagné. 2012. DEAP: Evolutionary algorithms made easy. Journal of Machine Learning Research 13, Jul (2012), 2171–2175

    2012

  3. [11]

    Fred Glover, Manuel Laguna, and Rafael Martí. 2003. Scatter search. In Advances in evolutionary computing. Springer, 519–537

    2003

  4. [12]

    Fred W Glover and Gary A Kochenberger. 2006. Handbook of metaheuristics . Vol. 57. Springer Science & Business Media

    2006

  5. [13]

    Goldberg

    David E. Goldberg. 1989. Genetic Algorithms in Search, Optimization and Machine Learning (1st ed.). Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA

    1989

  6. [14]

    Albert Gordo, Jon Almazán, Jerome Revaud, and Diane Larlus. 2016. Deep image retrieval: Learning global representations for image search. In European Conference on Computer Vision . Springer, 241–257

    2016

  7. [15]

    Albert Gordo, Jon Almazan, Jerome Revaud, and Diane Larlus. 2017. End-to-end learning of deep visual representations for image retrieval. International Journal of Computer Vision 124, 2 (2017), 237–254

    2017

  8. [16]

    John J Grefenstette. 1986. Optimization of control parameters for genetic algo- rithms. IEEE Transactions on systems, man, and cybernetics 16, 1 (1986), 122–128

    1986

  9. [17]

    Walter J. Gutjahr. 2010. Convergence Analysis of Metaheuristics . Springer US, Boston, MA, 159–187. https://doi.org/10.1007/978-1-4419-1306-7_6

    2010 doi

  10. [18]

    Holger H Hoos. 2011. Automated algorithm configuration and parameter tuning. In Autonomous search. Springer, 37–71

    2011

  11. [19]

    Mark J Huiskes and Michael S Lew. 2008. The MIR flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia Information Retrieval. ACM, 39–43

    2008

  12. [20]

    Piotr Indyk and Rajeev Motwani. 1998. Approximate nearest neighbors: towards removing the curse of dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing . ACM, 604–613

    1998

  13. [21]

    Ahmet Iscen, Giorgos Tolias, Yannis S Avrithis, Teddy Furon, and Ondrej Chum

  14. [22]

    Yannis Kalantidis, Clayton Mellina, and Simon Osindero. 2016. Cross-dimensional weighting for aggregated deep convolutional features. In European Conference on Computer Vision. Springer, 685–701

    2016

  15. [23]

    Giorgos Karafotias, Mark Hoogendoorn, and Ágoston E Eiben. 2015. Parameter Control in Evolutionary Algorithms: Trends and Challenges. IEEE Transactions on Evolutionary Computation 19, 2 (2015), 167–187

    2015

  16. [24]

    Scott Kirkpatrick, C Daniel Gelatt, and Mario P Vecchi. 1983. Optimization by simulated annealing. science 220, 4598 (1983), 671–680

    1983

  17. [25]

    SV Konstantinov, AI Diveev, GI Balandina, and AA Baryshnikov. 2019. Compara- tive Research of Random Search Algorithms and Evolutionary Algorithms for the Optimal Control Problem of the Mobile Robot. Procedia Computer Science 150 (2019), 462–470

    2019

  18. [26]

    Federico Magliani, Tomaso Fontanini, and Andrea Prati. 2019. Landmark Recogni- tion: From Small-Scale to Large-Scale Retrieval. In Recent Advances in Computer Vision. Springer, 237–259

    2019

  19. [27]

    Federico Magliani, Kevin McGuiness, Eva Mohedano, and Andrea Prati. 2019. An Efficient Approximate kNN Graph Method for Diffusion on Image Retrieval. arXiv preprint arXiv:1904.08668 (2019)

    2019 arXiv

  20. [28]

    Federico Magliani and Andrea Prati. 2018. An accurate retrieval through R-MAC+ descriptors for landmark recognition. In Proceedings of the 12th International Conference on Distributed Smart Cameras . ACM, 6

    2018

  21. [29]

    Pablo Mesejo, Oscar Ibáñez, Oscar Cordón, and Stefano Cagnoni. 2016. A survey on image segmentation using metaheuristic-based deformable models: state of the art and critical analysis. Applied Soft Computing 44 (2016), 1–29

    2016

  22. [30]

    RV Mises and Hilda Pollaczek-Geiringer. 1929. Praktische Verfahren der Gleichungsauflösung. ZAMM-Journal of Applied Mathematics and Mechan- ics/Zeitschrift für Angewandte Mathematik und Mechanik 9, 2 (1929), 152–164

    1929

  23. [31]

    Elizabeth Montero, María-Cristina Riff, and Nicolás Rojas-Morales. 2018. Tuners review: How crucial are set-up values to find effective parameter values? Engi- neering Applications of Artificial Intelligence 76 (2018), 108–118

    2018

  24. [32]

    Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank citation ranking: Bringing order to the web. Technical Report. Stanford InfoLab

    1999

  25. [33]

    Philbin, O

    J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman. 2007. Object Retrieval with Large Vocabularies and Fast Spatial Matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    2007

  26. [34]

    James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman

  27. [35]

    Riccardo Poli, James Kennedy, and Tim Blackwell. 2007. Particle swarm opti- mization. Swarm intelligence 1, 1 (2007), 33–57

    2007

  28. [36]

    Jussi Rasku, Nysret Musliu, and Tommi Kärkkäinen. 2019. On automatic algorithm configuration of vehicle routing problem solvers. Journal on Vehicle Routing Algorithms (22 Feb 2019). https://doi.org/10.1007/s41604-019-00010-9

    2019 doi

  29. [37]

    Moshe Sipper, Weixuan Fu, Karuna Ahuja, and Jason H Moore. 2018. Investigating the parameter space of evolutionary algorithms. BioData Mining 11, 1 (2018), 2

    2018

  30. [38]

    Roberto Ugolotti, Laura Sani, and Stefano Cagnoni. 2019. What Can We Learn from Multi-Objective Meta-Optimization of Evolutionary Algorithms in Contin- uous Domains? Mathematics 7, 3 (2019), 232

    2019

  31. [39]

    Fan Yang, Ryota Hinami, Yusuke Matsui, Steven Ly, and Shin’ichi Satoh. 2018. Efficient Image Retrieval via Decoupling Diffusion into Online and Offline Pro- cessing. arXiv preprint arXiv:1811.10907 (2018)

    2018 arXiv

  32. [40]

    Denny Zhou, Jason Weston, Arthur Gretton, Olivier Bousquet, and Bernhard Schölkopf. 2004. Ranking on data manifolds. In Advances in Neural Information Processing Systems. 169–176

    2004

  33. [2008]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Lost in quantization: Improving particular object retrieval in large scale image databases. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 1–8

  34. [2017]

    In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , Vol

    Efficient Diffusion on Region Manifolds: Recovering Small Objects with Compact CNN Representations.. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , Vol. 1. 3

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.