Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Prototype-Based Semantic Consistency Alignment for Domain Adaptive Retrieval

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A prototype-based two-stage hashing framework claims consistent gains in domain adaptive retrieval by aligning class semantics, weighting pseudo-labels by geometric agreement, and quantizing reconstructed rather than raw features.

desk verdict PSCA is a promising DAR hashing recipe with strong reported numbers and a genuinely reusable reconstruction idea, but the appendix's orthogonal-prototype derivation and dimension bookkeeping need correction before the theory can be trusted. read the letter →

arxiv 2512.04524 v4 pith:BBSJIERE submitted 2025-12-04 cs.LG cs.AI

Tianle Hu , Weijun Lv , Na Han , Xiaozhao Fang , Jie Wen , Jiaxing Li , Guoxu Zhou This is my paper
classification cs.LGcs.AI
keywords domainadaptiveretrievalhashingorthogonalprototypelearningpseudo-labelreliabilitysemanticconsistencyalignmentfeaturereconstructioncross-domainbinaryhashcodes
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

Domain adaptive retrieval — learning binary hash codes that work when queries and database come from different visual domains — has been held back, this paper argues, because methods align samples pairwise instead of classes, trust potentially wrong pseudo-labels, and hash features still contaminated by domain shift. PSCA tries to fix all three at once: it learns a set of mutually perpendicular class prototypes in a shared subspace, weights each target sample's pseudo-label by whether geometric proximity and semantic prediction agree, and reconstructs features as confidence-weighted combinations of prototypes before quantizing them into hash codes. The claimed payoff is consistent gains over existing methods across four benchmarks and multiple code lengths, with mean average precision improvements up to 17.21 percentage points on MNIST→USPS and an average 8.82 percentage-point gain on Office-Home. If true, the paper would establish class-level prototype alignment with geometric pseudo-label correction as a credible alternative to pairwise sample alignment in non-deep domain adaptive hashing.

What carries the argument

The load-bearing object is the set of class prototypes O ∈ R^{q×c}, required to be mutually orthogonal (OᵀO = I_c), which directly models class-level semantic structure in a domain-shared subspace. A soft membership matrix R assigns target samples to prototypes under a geometric-proximity term and a semantic-aware term, with a per-sample weight α that increases trust in pseudo-labels when geometry and semantics agree and decreases it when they conflict. The reconstructed feature matrix D concatenates prototype-weighted reconstructions with the projected original features, and stage-two hashing learns two domain-specific quantization functions Ws and Wt whose outputs are forced to approximate

What would settle it

Run the stage-one update on any reported benchmark and compute OᵀO. If off-diagonal entries are materially nonzero, the prototypes are not orthogonal. A second check: replace the SVD update with an explicitly orthogonal update (e.g., polar decomposition or projected gradient); if mean average precision stays the same, orthogonal prototypes are not the active ingredient behind the reported gains.

Watch

Extended reading notes

Core claim

On the paper's own terms, its central discovery is that the limiting factor in domain adaptive retrieval is not just distribution gap but the way semantics are used. By replacing pairwise sample alignment with orthogonal class prototypes, PSCA aligns classes directly and forces inter-class separation; by blending geometric proximity with pseudo-label predictions, it produces a soft membership matrix that corrects unreliable labels; and by reconstructing features from prototypes and memberships before hashing, it avoids quantizing domain-shifted raw features. The result is a unified Hamming space learned through two domain-specific quantizers that are pulled toward each other, and the paper r

Load-bearing premise

The load-bearing premise is that the learned class centers remain mutually perpendicular after each update; the derivation only offers an approximate SVD-based solution, and if the prototypes are not actually orthogonal, the claimed inter-class separation and the semantic-alignment justification weaken.

Editorial extensions

If this is right

  • Class-level prototype alignment can replace expensive pairwise sample comparisons, which the paper argues reduces computational cost and covers distributional structure more broadly.
  • Geometric proximity is a usable, continuous reliability signal for pseudo-labels: the adaptive weighting scheme automatically down-weights assignments when geometry and semantics disagree.
  • Quantizing reconstructed, prototype-enhanced features instead of raw projected features reduces the harm of domain shift and improves hash code quality before any code is produced.
  • Two domain-specific quantizers with a mutual approximation constraint can generate a unified Hamming space while still capturing domain-specific characteristics.
  • The reported margins — up to 17.21% MAP over the best prior non-deep method on MNIST→USPS and an 8.82% average on Office-Home — imply the approach scales across small, medium, and large benchmarks.

Reading between the lines

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

  • Inference: The theoretical hinge is the orthogonality of prototypes, yet the closed-form SVD update O = M̂ Mᵀ does not guarantee OᵀO = I_c; a reader could compute the residual and see whether the separation claim holds exactly or only approximately.
  • Inference: The smaller single-domain gains (2.25% on A→D) and the paper's own over-smoothing comment suggest the reconstruction step should be made adaptive — for example, by re-weighting the geometric and reconstructed feature blocks per domain — which the paper does not explore.
  • Inference: Because every component is a linear projection or prototype combination, the scheme is a natural plug-in regularizer for deep domain adaptive retrieval pipelines, not just a standalone shallow hashing method.
  • Inference: A sharper test of the pseudo-label correction story would be to measure performance as a function of label noise rate; the current experiments use real benchmarks where noise is uncontrolled.
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 PSCA, a two-stage non-deep method for domain adaptive retrieval (DAR) hashing. Stage one learns a projection P and class prototypes O in a shared subspace under an MMD term, an l2,1 sparsity term, and an orthogonality constraint O^T O = I_c. Pseudo-labels are generated by Nearest Class Prototype and Structured Prediction, then corrected by an adaptive semantic-consistency weight alpha based on geometric proximity, yielding a soft membership matrix R. Stage one reconstructs features by concatenating prototype-weighted representations with projected features. Stage two learns two domain-specific linear quantizers W_s, W_t under mutual approximation and orthogonality constraints, producing unified binary codes via discrete cyclic coordinate descent. Experiments on Office-31, Office-Home, COIL20, and MNIST-USPS compare against 11 shallow/transfer hashing baselines and 3 deep baselines, with MAP, precision-recall, Top-K, ablation, parameter sensitivity, convergence, and runtime analyses.

Significance. PSCA targets a genuine limitation of existing shallow DAR hashing methods: pairwise alignment without class-level semantic structure, unreliable pseudo-labels, and premature quantization of domain-shifted features. The proposed pipeline is well motivated, and the experimental design is broad: four benchmark families, multiple code lengths, single- and cross-domain settings, comparison with recent deep methods, ablations, and parameter sensitivity studies. If the orthogonal-prototype derivation is corrected and the performance is reproducible, PSCA would be a strong non-deep DAR baseline and a useful reference point for the field. However, the central theoretical derivation of the key novel component is not sound as written, and the empirical SOTA claim is weakened by the absence of variance reporting and code. The significance is therefore conditional on a substantial revision of the optimization appendix.

major comments (4)
  1. [Appendix, Eqs. (19), (26)-(28)] The orthogonal-prototype update is not derived correctly. In Eq. (19), the expansion is written as a single trace Tr(P^T X S1 X^T P + O^T S2 O - 2 P^T X \tilde Y O^T), but the three terms are q x q, c x c, and q x q; the expression is undefined unless q = c, while the complexity section states c < q. If the intended form is a sum of traces, it should be written as such. More importantly, the SVD update in Eqs. (27)-(28) does not solve the constrained problem in Eq. (26). Setting the derivative to zero ignores the constraint O^T O = I_c. For a thin SVD A = \hat M \Gamma M^T with A = P^T X \tilde Y S2^{-1}, the matrix O = \hat M M^T satisfies O^T O = M M^T, which equals I_c only when k = c and M is square orthogonal. The paper neither states that the full, untruncated SVD is used nor reports the resulting orthogonality error. Since Eq. (3) has no feasible solution when q < c, and q is not
  2. [Main text Eq. (7); Appendix Eq. (15); Algorithm 1 line 11] The stage-two orthogonality constraints are inconsistent. The main text imposes W_s W_s^T = I_r and W_t W_t^T = I_r (orthonormal rows), while Appendix Eq. (15) imposes W_s^T W_s = I_C and W_t^T W_t = I_C (orthonormal columns). Since W_s is r x C with r < C, the latter is infeasible by rank. The SVD update in Eqs. (33)-(34) yields W_s = \hat S S^T, which satisfies W_s W_s^T = I_r only when \hat S is square and orthonormal, so the main-text constraint is the workable one; the appendix must be corrected to match. In addition, Algorithm 1 line 11 defines C = q + d for D = [\tilde X, X^T P]^T, whereas the main text defines C = 2q. This changes the dimensions of W_s, W_t and the complexity estimates. Please reconcile these definitions and report the actual C used.
  3. [Tables 1-3] The empirical SOTA claim is not fully supported by the reported statistics. The text states that each trial is repeated 10 times, but Tables 1-3 report only the mean MAP with no standard deviation, confidence interval, or significance test. Some of the reported advantages over the second-best baseline are small (e.g., Table 1, A->D at 16 bits: 56.44 vs 55.13; COIL1->COIL2 at 16 bits: 84.74 vs 82.44), and without variance information the reader cannot assess whether the gains are stable across repetitions. Since the paper's central claim is consistent superiority, I ask the authors to report standard deviations (or equivalent) and, ideally, release code to make the experiments reproducible.
  4. [Appendix Lemma 1 and Eq. (23)] The convergence analysis is not a proof for the actual updates. Lemma 1 states that Algorithm 1 monotonically decreases L(W) conditional on all subproblems reaching their global optima, but the O update in Eq. (28) is not the exact solution of the constrained subproblem, and the R update in Eq. (24) is a single gradient step followed by a projection, not a global optimization. Thus the monotone-descent property is not established for the implemented procedure. The empirical convergence curves are useful, but should be presented as empirical evidence, not as a formal theorem. Also, Eq. (23) contains an unexplained ln(2) in the derivative of -psi_ij log(r_ij); with a natural logarithm the derivative is -psi_ij / r_ij. Please clarify or correct.
minor comments (5)
  1. [Eq. (5)] The definitions k_geo = arg min_j d_ij and k_sem = arg max_j pi_ij do not specify tie-breaking. Ties are unlikely but possible, especially at initialization; a deterministic tie rule would avoid ambiguity.
  2. [Algorithm 1, lines 9-10] The comments for reconstructing source and target features are both labeled 'Reconstruct source feature representation'; the second should read 'target'. Also, Algorithm 1 line 11's C = q + d conflicts with the main text; see major comment 2.
  3. [Table 2] The PWCF row for R->C is non-monotonic in code length: 12.79 at 16 bits, 21.97 at 64 bits, but 10.39 at 128 bits. Please check whether this is a typo and clarify in the text.
  4. [Figure 1] The framework figure is difficult to read: many labels are rendered with garbled or missing glyphs, and the legend for the semantic consistency alignment panels is not legible at normal print size. Please regenerate with vector fonts and larger labels.
  5. [General] Minor typos include 'indexs' near Eq. (5), 'opitmize' in the Appendix, and inconsistent use of 'source feature representation' in Algorithm 1. A careful proofread would improve clarity.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: PSCA is evaluated on external benchmarks and its prototype/hash pipeline is feed-forward; the only self-citations are background, and the supplementary SVD issue is a correctness gap, not a circular reduction.

full rationale

Walking the derivation chain: pseudo-labels (Eqs. 10-13) initialize the soft membership R (Eq. 4) with confidence weighting α (Eq. 5); R and O are then used to reconstruct features (D = [X̃, X^T P]^T), and stage-two hashing (Eq. 7) maps these to binary codes. Every stage is a forward optimization with measured MAP on public benchmarks (Office-31, Office-Home, COIL20, MNIST-USPS); no parameter is fitted to a target MAP value and then reported as a prediction. The citations to the authors' own earlier work, 'Hu et al. 2025a' and 'Hu et al. 2025b', occur only as background ('To bridge this non-negligible domain gap (Hu et al. 2025a)'; 'Hashing ... (Hu et al. 2025b)') and do not carry the load of the PSCA derivation or the SOTA claim. The genuine weakness in the supplementary material is not circularity: Eq. (19) writes Tr(O^T S2 O) with O ∈ R^{q×c}, S2 ∈ R^{c×c}, which is only defined if q=c; and the SVD relaxation in Eqs. (27)-(28), O = M̂ M^T, is not shown to satisfy the constraint O^T O=I_c of Eq. (14). That means the 'orthogonal prototype' claim is under-supported mathematically, but it is an optimization/derivation gap in the method, not a reduction of the reported results to the inputs. The paper also honestly notes its A→D single-domain limitation ('PSCA shows the least performance improvement on A→D by 2.25%'), which is a stated limitation, not a circular step. Overall, the central empirical claim is not forced by construction: the results stand on external MAP comparisons and standard metrics.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. The 'soft membership matrix' and 'domain-specific quantization functions' are algorithmic constructs, not externally evidenced entities. The main free parameters are the tuning hyperparameters; ideally these would be fixed a priori, but they are data-dependent. The key assumptions are standard for the DAR literature but are not independently validated in this paper.

free parameters (6)
  • lambda_1 (MMD weight)
    Tuned per dataset from {1,10,100}; controls degree of marginal alignment.
  • lambda_2 (sparsity weight)
    Tuned per dataset from {0.1,1,10}; controls row-sparsity of P.
  • lambda_3 (quantization agreement)
    Tuned per dataset from {1,10,100}; controls similarity between W_s and W_t.
  • sigma (nonlinear down-weighting exponent) = 2
    Manually set; amplifies penalty differences for membership R.
  • beta (regularization on Phi) = 0.1
    Fixed constant in out-of-sample mapping.
  • subspace dimension q
    Not specified in the text; must be chosen by the user and affects all subsequent computations.
assumptions (4)
  • domain assumption MMD alignment on projected features reduces cross-domain marginal discrepancy.
    Used in Eq. (1) and throughout; assumes that aligning means in a shared subspace is beneficial for retrieval.
  • domain assumption Pseudo-labels from NCP and SP are sufficiently informative to guide prototype learning.
    Pseudo-labels are generated before P is fully learned; their quality is not quantified, and the method relies on them to seed the semantic consistency alignment.
  • domain assumption The alternating optimization of nonconvex objectives with closed-form subproblem solutions reaches a stationary point.
    Lemma 1 claims monotonic decrease but the argument is informal; the SVD-based updates are relaxations that may not decrease the original objective.
  • domain assumption Reconstructed features D = [\tilde{X}, X^T P] retain enough information for hash learning.
    The paper assumes that the concatenation of prototype reconstructions and projected features preserves both semantic and geometric structure; no theoretical or empirical justification beyond ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prototype-Based Semantic Consistency Alignment for Domain Adaptive Retrieval." pith.science (2026). https://pith.science/paper/BBSJIERE

@misc{pith2026251204524,
  author       = {Pith},
  title        = {Pith review of: Prototype-Based Semantic Consistency Alignment for Domain Adaptive Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBSJIERE}},
  note         = {Machine review of arXiv:2512.04524}
}
read the original abstract

Domain adaptive retrieval aims to transfer knowledge from a labeled source domain to an unlabeled target domain, enabling effective retrieval while mitigating domain discrepancies. However, existing methods encounter several fundamental limitations: 1) neglecting class-level semantic alignment and excessively pursuing pair-wise sample alignment; 2) lacking either pseudo-label reliability consideration or geometric guidance for assessing label correctness; 3) directly quantizing original features affected by domain shift, undermining the quality of learned hash codes. In view of these limitations, we propose Prototype-Based Semantic Consistency Alignment (PSCA), a two-stage framework for effective domain adaptive retrieval. In the first stage, a set of orthogonal prototypes directly establishes class-level semantic connections, maximizing inter-class separability while gathering intra-class samples. During the prototype learning, geometric proximity provides a reliability indicator for semantic consistency alignment through adaptive weighting of pseudo-label confidences. The resulting membership matrix and prototypes facilitate feature reconstruction, ensuring quantization on reconstructed rather than original features, thereby improving subsequent hash coding quality and seamlessly connecting both stages. In the second stage, domain-specific quantization functions process the reconstructed features under mutual approximation constraints, generating unified binary hash codes across domains. Extensive experiments validate PSCA's superior performance across multiple datasets.

Figures

Figures reproduced from arXiv: 2512.04524 by the authors.

Figure 1
Figure 1. The framework of the proposed PSCA Nearest Class Prototype (NCP) and Structured Prediction (SP) (Wang and Breckon 2020), where c is the number of categories. Let π (i,e) sort be the e-th largest element in πi , where π (i,1) sort ≥ π (i,2) sort ≥ · · · ≥ π (i,c) sort . Then the pseudo-label is assigned according to π (i,1) sort . The detailed pseudo-labeling process is presented in the supplementary materials. Defin… view at source ↗
Figure 2
Figure 2. The Precision-Recall Curves and Top-K Precision [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparison with deep methods on MNIST-USPS [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Distribution visualization on case MNIST [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Convergence analysis of PSCA. jective function value L(W). On the basis of the Bounded Monotone Convergence theory (Russell 2020), Algorithm 1 achieves good convergence property of PSCA.  For visual analysis, we plot convergence curves on all benchmark datasets. As il…
Figure 6
Figure 6. Figure 6: Parameter sensitivity analysis of PSCA on cases [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Chen, C.; Xie, W.; Huang, W.; Rong, Y.; Ding, X.; Huang, Y.; Xu, T.; and Huang, J. 2019. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 627--636

    2019

  4. [4]

    Chen, Y.; Fang, X.; Liu, Y.; Zheng, W.; Kang, P.; Han, N.; and Xie, S. 2023. Two-step strategy for domain adaptation retrieval. IEEE Transactions on Knowledge and Data Engineering, 36(2): 897--912

    2023

  5. [5]

    Cui, H.; Zhao, L.; Li, F.; Zhu, L.; Han, X.; and Li, J. 2024. Effective Comparative Prototype Hashing for Unsupervised Domain Adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 8329--8337

    2024

  6. [6]

    Feng, L.; Huang, J.; Shu, S.; and An, B. 2020. Regularized matrix factorization for multilabel learning with missing labels. IEEE transactions on cybernetics, 52(5): 3710--3721

    2020

  7. [7]

    Gong, Y.; Lazebnik, S.; Gordo, A.; and Perronnin, F. 2012. Iterative quantization: A procrustean approach to learning binary codes for large-scale image retrieval. IEEE transactions on pattern analysis and machine intelligence, 35(12): 2916--2929

    2012

  8. [8]

    M.; Rasch, M

    Gretton, A.; Borgwardt, K. M.; Rasch, M. J.; Sch \"o lkopf, B.; and Smola, A. 2012. A kernel two-sample test. The Journal of Machine Learning Research, 13(1): 723--773

    2012

Show all 42 references
  1. [9]

    Hu, T.; Chen, Y.; Cheng, C.; Xiao, J.; Sun, W.; and Fang, X. 2025 a . Coarse-to-Fine Label Refinement for Domain Adaptive Retrieval. Information Sciences, 122532

    2025

  2. [10]

    Hu, T.; Chen, Y.; Lv, W.; Chen, Y.; and Fang, X. 2025 b . Consistent coding guided domain adaptation retrieval. Applied Intelligence, 55(7): 706

    2025

  3. [11]

    Huang, F.; Zhang, L.; and Gao, X. 2021. Domain adaptation preconceived hashing for unconstrained visual retrieval. IEEE Transactions on Neural Networks and Learning Systems, 33(10): 5641--5655

    2021

  4. [12]

    Huang, F.; Zhang, L.; Yang, Y.; and Zhou, X. 2020. Probability weighted compact feature for domain adaptive retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9582--9591

    2020

  5. [13]

    Hull, J. J. 1994. A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence, 16(5): 550--554

    1994

  6. [14]

    Jiang, Q.-Y.; and Li, W.-J. 2015. Scalable graph hashing with feature transformation. In IJCAI, volume 15, 2248--2254

    2015

  7. [15]

    LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11): 2278--2324

    1998

  8. [16]

    Liu, H.; Ji, R.; Wang, J.; and Shen, C. 2018. Ordinal constraint binary coding for approximate nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 41(4): 941--955

    2018

  9. [17]

    Liu, H.; Wang, R.; Shan, S.; and Chen, X. 2016. Deep supervised hashing for fast image retrieval. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2064--2072

    2016

  10. [18]

    Liu, X.; Li, J.; Nie, X.; Zhang, X.; and Yin, Y. 2024. Fast Unsupervised Cross-Modal Hashing with Robust Factorization and Dual Projection. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12): 1--21

    2024

  11. [19]

    Long, M.; Wang, J.; Ding, G.; Sun, J.; and Yu, P. S. 2013. Transfer feature learning with joint distribution adaptation. In Proceedings of the IEEE international conference on computer vision, 2200--2207

    2013

  12. [20]

    Long, M.; Wang, J.; Ding, G.; Sun, J.; and Yu, P. S. 2014. Transfer joint matching for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1410--1417

    2014

  13. [21]

    Luo, J.; Zhao, Y.; Luo, X.; Xiao, Z.; Ju, W.; Shen, L.; Tao, D.; and Zhang, M. 2025. Cross-domain diffusion with progressive alignment for efficient adaptive retrieval. IEEE Transactions on Image Processing

    2025

  14. [22]

    Luo, X.; Zhang, P.-F.; Huang, Z.; Nie, L.; and Xu, X.-S. 2019. Discrete hashing with multiple supervision. IEEE Transactions on Image Processing, 28(6): 2962--2975

    2019

  15. [23]

    A.; Nayar, S

    Nene, S. A.; Nayar, S. K.; Murase, H.; et al. 1996. Columbia object image library (coil-20)

    1996

  16. [24]

    Nie, F.; Huang, H.; Cai, X.; and Ding, C. 2010. Efficient and robust feature selection via joint 2, 1-norms minimization. Advances in neural information processing systems, 23

    2010

  17. [25]

    Russell, B. 2020. Principles of mathematics. Routledge

    2020

  18. [26]

    Saenko, K.; Kulis, B.; Fritz, M.; and Darrell, T. 2010. Adapting visual category models to new domains. In Computer Vision--ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, 213--226. Springer

    2010

  19. [27]

    Shen, F.; Shen, C.; Liu, W.; and Tao Shen, H. 2015. Supervised discrete hashing. In Proceedings of the IEEE conference on computer vision and pattern recognition, 37--45

    2015

  20. [28]

    Venkateswara, H.; Eusebio, J.; Chakraborty, S.; and Panchanathan, S. 2017. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5018--5027

    2017

  21. [29]

    Wang, H.; Sun, J.; Luo, X.; Xiang, W.; Zhang, S.; Chen, C.; and Hua, X.-S. 2023 a . Toward effective domain adaptive retrieval. IEEE Transactions on Image Processing, 32: 1285--1299

    2023

  22. [30]

    Wang, Q.; and Breckon, T. 2020. Unsupervised domain adaptation via structured prediction based selective pseudo-labeling. In Proceedings of the AAAI conference on artificial intelligence, volume 34, 6243--6250

    2020

  23. [31]

    Wang, Q.; and Chen, K. 2017. Zero-shot visual recognition via bidirectional latent embedding. International Journal of Computer Vision, 124: 356--383

    2017

  24. [32]

    Wang, S.; Liu, X.; Zhu, E.; Tang, C.; Liu, J.; Hu, J.; Xia, J.; and Yin, J. 2019. Multi-view clustering via late fusion alignment maximization. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, 3778--3784

    2019

  25. [33]

    Wang, Z.; Xiao, H.; Duan, Y.; Zhou, J.; and Lu, J. 2023 b . Learning Deep Binary Descriptors via Bitwise Interaction Mining. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2): 1919--1933

    2023

  26. [34]

    Weiss, Y.; Torralba, A.; and Fergus, R. 2008. Spectral hashing. Advances in neural information processing systems, 21

    2008

  27. [35]

    E.; Jie, Z.; Liu, W.; and Feng, J

    Yuan, L.; Wang, T.; Zhang, X.; Tay, F. E.; Jie, Z.; Liu, W.; and Feng, J. 2020. Central similarity quantization for efficient image and video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3083--3092

    2020

  28. [36]

    Zhang, L.; Liu, J.; Yang, Y.; Huang, F.; Nie, F.; and Zhang, D. 2019. Optimal projection guided transfer hashing for image retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 30(10): 3788--3802

    2019

  29. [37]

    Zhang, W.; Yang, X.; Teng, S.; and Wu, N. 2023 a . Semantic-guided hashing learning for domain adaptive retrieval. World Wide Web, 26(3): 1093--1112

    2023

  30. [38]

    Zhang, W.; Zhou, K.; Teng, L.; Tang, F.; Wu, N.; Teng, S.; and Li, J. 2024. Dynamic Confidence Sampling and Label Semantic Guidance Learning for Domain Adaptive Retrieval. IEEE Transactions on Multimedia, 26: 2467--2479

    2024

  31. [39]

    Zhang, Y.; Tian, S.; Liao, M.; Zhang, Z.; Zou, W.; and Xu, C. 2023 b . Fine-grained self-supervision for generalizable semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1): 371--383

    2023

  32. [40]

    Zhang, Y.; Tian, S.; Liao, M.; Zou, W.; and Xu, C. 2023 c . A hybrid domain learning framework for unsupervised semantic segmentation. Neurocomputing, 516: 133--145

    2023

  33. [41]

    T.; Zhao, H.; Peng, X.; Fang, M.; Qin, Z.; and Goh, R

    Zhou, J. T.; Zhao, H.; Peng, X.; Fang, M.; Qin, Z.; and Goh, R. S. M. 2018. Transfer hashing: From shallow to deep. IEEE transactions on neural networks and learning systems, 29(12): 6191--6201

    2018

  34. [42]

    Zhu, L.; Lu, X.; Cheng, Z.; Li, J.; and Zhang, H. 2020. Deep collaborative multi-view hashing for large-scale image search. IEEE Transactions on Image Processing, 29: 4643--4655

    2020

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.