Pith. sign in

REVIEW 5 major objections 5 minor 34 references

FAR-Net: Multi-Stage Fusion Network with Enhanced Semantic Alignment and Adaptive Reconciliation for Composed Image Retrieval

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

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

Pith's one-line read FAR-Net boosts composed image retrieval by up to 2.4 percent Recall@1 by cascading early and late fusion.

desk verdict Incremental CIR architecture with a plausibly useful late-to-early cascade, but the contrastive losses as written are mathematically broken and the SOTA comparison is uncontrolled. read the letter →

arxiv 2507.12823 v1 pith:PLE5O22T submitted 2025-07-17 cs.CV

Jeong-Woo Park , Young-Eun Kim , Seong-Whan Lee This is my paper
classification cs.CV
keywords composedimageretrievalearlyfusionlateQ-Formercross-attentionuncertaintyembeddingFashionIQCIRRcontrastivelearning
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

Composed image retrieval finds a target image from a reference photo plus a text description of the desired change. The paper argues that the two standard fusion strategies, early fusion and late fusion, are not an either-or choice and can be combined in one cascade. It proposes FAR-Net, whose ESAM stage does late fusion with cross-attention for fine-grained text-to-region alignment, and whose ARM stage applies early fusion with uncertainty-scaled perturbations for robustness. On the FashionIQ and CIRR benchmarks, it reports consistent gains over previous methods, up to 2.4% in Recall@1 and 1.0% in Recall@50. If the result holds, hybrid fusion is a viable route to more accurate user-specified image search.

What carries the argument

The central object is FAR-Net's two-module cascade: ESAM (enhanced semantic alignment module) carries late fusion with a Q-Former cross-attention map that aligns textual tokens to image regions, while ARM (adaptive reconciliation module) carries early fusion with uncertainty embeddings that perturb the target representation. The attention map $A_i$ is the mechanism that transfers fine-grained semantic alignment between the reference and target branches, and the uncertainty-scaled perturbation $\hat{v}_i$ is the mechanism that prevents the model from over-relying on explicitly mentioned textual cues. Together they make the total loss $L_{Total}$ enforce global alignment, token-region consistency, robustness to perturbation, and prompt-target fidelity in a single training pipeline.

What would settle it

Retrain each leading baseline with the same ViT-L image encoder, 224-by-224 input size, batch size of 32, 50 epochs, and AdamW settings as FAR-Net, then run the same evaluation split; if the Recall@1 gap on CIRR falls below the reported 2.4 points, the fusion modules are not what drive the advantage.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Late-to-Early cascaded architecture improves composed image retrieval beyond what either fusion paradigm achieves alone. ESAM encodes reference image and modification text separately, fuses them as $F_{ui} = \lambda_1 f_\phi(R_i) + (1-\lambda_1) f_\zeta(M_i)$, and optimizes a contrastive loss $L_{Late}$; a Q-Former computes cross-attention maps $A_i = \mathrm{softmax}(Q_i K_i^\top / \sqrt{d})$ for reference-modification and target-modification pairs, and an attention-alignment loss pushes the two maps together. ARM then re-optimizes the fused representation with perturbed target embeddings $\hat{v}_i = \alpha v_i + \beta$, where $\alpha \sim \mathcal{N}(1, \sigma_t)$ and $\beta \sim \mathcal{N}(\mu_t, \sigma_t)$, under a resilience loss that combines deterministic and uncertain embeddings, plus a prompt-to-image contrastive loss that aligns the text prompt with the target image. The total objective $L_{Total} = L_{ESAM} + L_{ARM}$ is trained end-to-end, and the paper reports state-of-the-art Recall@1 of 54.39 on CIRR and the best average retrieval numbers on FashionIQ among the compared methods.

Load-bearing premise

The load-bearing premise is that the reported benchmark gains come from the ESAM and ARM fusion modules, not from FAR-Net's stronger ViT-L encoder or from differences in training recipe compared with the baselines.

Editorial extensions

If this is right

  • On CIRR, the reported Recall@1 of 54.39 means that for more than half of text-modified queries, the correct image is the model's first retrieval; the Rsubset@1 of 80.77 means the same is true for four fifths of queries from the visually similar subset.
  • The ablation study attributes the largest single drop to removing the late-fusion contrastive loss (18.96 points on FashionIQ), indicating that global alignment is the strongest component, with attention alignment, prompt-to-image loss, and resilience loss each contributing smaller but measurable gains.
  • Because the full pipeline is trained on standard triplets of reference image, modification text, and target image, the same architecture can be applied to any new dataset that supplies this triplet format, without extra annotations.
  • The reported results support the paper's structural claim that early and late fusion are complementary: a cascade can keep modality-specific features while still obtaining token-level cross-modal interactions.
  • The consistent gains on both a fashion-specific and an open-domain benchmark suggest that hybrid fusion generalizes beyond a single retrieval domain.

Reading between the lines

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

  • The paper leaves implicit that a backbone-matched comparison is needed: FAR-Net uses a ViT-L image encoder, while the comparison tables do not state the backbones used by the baselines, so part of the reported 2.4-point gap could be an encoder effect rather than an effect of the ESAM and ARM modules.
  • A clean control experiment, in which the same ViT-L encoder and training recipe are given to a simple late-fusion baseline, would tell whether uncertainty perturbations alone reproduce ARM's gains; the paper does not run this control.
  • The attention-map alignment between a reference-modification pair and a target-modification pair is a general scoring mechanism that could be transferred to text-guided image editing evaluation, where the question is whether the edited region matches the textual instruction.
  • The uncertainty-perturbation scheme is defined independently of the fusion architecture, so it could be tested as a drop-in regularizer for other contrastive vision-language retrieval models.
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

5 major / 5 minor

Summary. The paper proposes FAR-Net, a two-stage composed image retrieval architecture. The first stage (ESAM) combines late fusion of reference image and modification text with a cross-attention-based fine-grained alignment objective; the second stage (ARM) applies early-fusion-style uncertainty perturbations and a prompt-to-image objective. Experiments on FashionIQ and CIRR report gains over published state-of-the-art methods of up to 2.4% in Recall@1 and 1.0% in Recall@50, and ablations show that each loss term contributes to the final average. The central claim is that the combination of ESAM and ARM yields state-of-the-art retrieval accuracy on these two benchmarks.

Significance. If the empirical claims survive correction of the loss definitions and a properly controlled comparison, the paper would offer a plausible hybrid of early and late fusion with an uncertainty-based robustness term, and the ablations suggest both modules matter. The paper ships no code, and the technical specification in Eqs. (4), (6), and (8) is not a valid contrastive objective as written; combined with the unmatched backbones in Tables I-II, the reported gains cannot currently be attributed to the proposed ESAM/ARM mechanisms. With the issues addressed, the architecture idea is worth testing, but as submitted the evidence does not establish the stated SOTA claim.

major comments (5)
  1. [Section III-B, Eq. (4)] As written, the denominator sums over the diagonal similarities of other batch elements rather than over negatives for the i-th anchor: it uses exp(cos(A_j^T, A'_j)/tau) and never involves A_i. This is not a contrastive loss because the off-diagonal pair (A_i, A'_j) for j != i never appears as a negative; if all diagonal similarities are equal, the gradient vanishes, so the objective cannot pull A_i toward A'_i. The same structural index error appears in Eqs. (6) and (8), where the denominators are sum_j exp(u_j^T v*_j / tau) and sum_j exp(u'_j^T v_j / tau), respectively. The authors must replace these denominators with anchor-conditioned sums over negatives, or state explicitly that the equations are typos and that the experiments used the standard anchor-conditioned form.
  2. [Section III-B, Eq. (4)] There is also a shape mismatch in the cosine similarity of Eq. (4). A_i = softmax(Q_i K_i^T / sqrt(d)) is a matrix of attention weights between image regions and text tokens (say R x T), while cos(A_i^T, A'_i) treats the transposed and untransposed maps as vectors. Unless both maps are flattened before taking the cosine, the dimensions are incompatible, and the paper does not state that they are flattened.
  3. [Section IV-C, Tables I-II] The state-of-the-art comparison is not backbone-controlled. Section IV-B specifies only FAR-Net's ViT-L image encoder and BLIP-2/BERT text encoder; the tables do not state the encoders, training splits, or evaluation protocols used for CASE, SPRC, PAIR, TG-CIR, MEDIAN, and the other baselines. Since CIR results are highly sensitive to backbone and protocol, the reported 2.43% R@1 gain on CIRR cannot be cleanly attributed to ESAM and ARM rather than to the choice of encoder or training recipe.
  4. [Section IV-B and Section III-C] Several hyperparameters and architectural details necessary for reproduction are missing: the values of lambda_1 and lambda_2 in Eqs. (1) and (7), the temperature tau, the statistics mu_t and sigma_t of the uncertainty perturbations, and the construction of the Q-Former queries Q_i and Q'_i (number of learnable queries, whether they are initialized from image features, and projection dimensions). Without these, the proposed method cannot be reimplemented from the manuscript alone.
  5. [Section IV-D, Table III] The ablation setup is under-specified. The row 'ARM only' (40.18 on FashionIQ) is not defined: does it retain the late-fusion global objective or the Q-Former branch, and what exactly is removed? In addition, no standard deviations or number of random seeds are reported for any row, so the claim of 'consistent performance gains' is not statistically supported.
minor comments (5)
  1. [Abstract and Section I] The abstract reports an improvement of 1.04% in Recall@50, but the numbers in Tables I-II show a maximum R@50 gap of 0.56 points on FashionIQ and 0.14 points on CIRR; the source of the 1.04% figure should be clarified.
  2. [Section IV-B] Citations [31] and [32] are unrelated to the ViT-L backbone (they concern handwritten-numeral recognition and contour tracking), and reference [9] also appears off-topic for composed image retrieval; these should be replaced with appropriate references.
  3. [Section III-C] The symbol U_i is introduced in the text without a definition; it should be connected to F_ui from Eq. (1) or defined separately to avoid confusion.
  4. [Figure 2] The text in Figure 2 is very small and difficult to read; a higher-resolution figure or a simplified schematic would improve clarity.
  5. [Table I] The column header 'Avg.' is ambiguous: the table appears to average R@10 and R@50 across the three FashionIQ categories, but this should be stated explicitly in the caption or text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the retrieval objectives are supervised by external benchmark targets and no load-bearing claim reduces to its own input.

full rationale

FAR-Net's derivation chain is not circular. The global alignment loss in Eq. (2) is a standard in-batch contrastive objective that uses the ground-truth target embedding v_i from the public benchmark, and the attention, resilience, and prompt-to-image losses in Eqs. (4)-(9) are additional supervised objectives defined on the same externally annotated triplets; none of these losses is fitted to the reported Recall@K metrics and then presented as a prediction. The tunable hyperparameters (lambda_1, lambda_2, temperature tau) are free parameters, not fitted quantities renamed as predictions. The self-citations ([9], [14], [15], [31], [32]) concern unrelated document-analysis, tracking, and pattern-recognition topics and are not load-bearing for the CIR architecture. No uniqueness theorem is imported from the authors' prior work, and the Q-Former and ViT-L/BLIP-2 backbones are external pretrained components whose use is not justified by a circular argument. Two non-circular concerns are worth noting explicitly: the empirical SOTA comparison in Tables I-II may have uncontrolled baselines since Section IV-B gives FAR-Net's backbone but not those of the baselines, and Eqs. (4), (6), and (8) write denominators that sum diagonal similarities without anchor-specific negatives, which is a technical correctness issue rather than a circularity. Neither concern makes the central claim equivalent by construction to its inputs, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a handful of unstated hyperparameters and a few domain assumptions: pretrained encoders combine cleanly, cross-attention maps can be matched, and Gaussian perturbations model retrieval variability. No new physical or conceptual entities are introduced. The biggest unstated burden is that the architecture details and training constants are not fully specified.

free parameters (4)
  • lambda1 = not reported
    Eq. (1) weights the reference image and text features in the fused representation; its value is never stated and no sensitivity analysis is given.
  • lambda2 = not reported
    Eq. (7) weights the deterministic and perturbed contrastive losses in the Resilience Loss; its value is never stated.
  • temperature tau = not reported
    Used in all contrastive losses (Eqs. 2, 4, 6, 8); the value is not specified.
  • uncertainty noise statistics mu_t, sigma_t = not reported
    Section III-C1 samples alpha ~ N(1, sigma_t) and beta ~ N(mu_t, sigma_t); how these statistics are computed from target features is unspecified.
assumptions (3)
  • domain assumption Pretrained ViT-L and BERT/BLIP-2 encoders produce compatible features for linear combination and contrastive learning.
    Invoked in Eq. (1) and throughout Section III; no fine-tuning or alignment verification is provided for the combined features.
  • ad hoc to paper Cross-attention maps over (reference, text) and (target, text) are comparable after cosine similarity in Eq. (4).
    The LAttention loss assumes token-region alignment maps can be directly matched across reference and target images; this is a training heuristic, not a derived equivalence.
  • ad hoc to paper Gaussian multiplicative and additive perturbations model real retrieval input variability.
    Section III-C1 introduces alpha and beta sampled from normal distributions; no evidence links these noise statistics to actual retrieval variability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FAR-Net: Multi-Stage Fusion Network with Enhanced Semantic Alignment and Adaptive Reconciliation for Composed Image Retrieval." pith.science (2026). https://pith.science/paper/PLE5O22T

@misc{pith2026250712823,
  author       = {Pith},
  title        = {Pith review of: FAR-Net: Multi-Stage Fusion Network with Enhanced Semantic Alignment and Adaptive Reconciliation for Composed Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PLE5O22T}},
  note         = {Machine review of arXiv:2507.12823}
}
read the original abstract

Composed image retrieval (CIR) is a vision language task that retrieves a target image using a reference image and modification text, enabling intuitive specification of desired changes. While effectively fusing visual and textual modalities is crucial, existing methods typically adopt either early or late fusion. Early fusion tends to excessively focus on explicitly mentioned textual details and neglect visual context, whereas late fusion struggles to capture fine-grained semantic alignments between image regions and textual tokens. To address these issues, we propose FAR-Net, a multi-stage fusion framework designed with enhanced semantic alignment and adaptive reconciliation, integrating two complementary modules. The enhanced semantic alignment module (ESAM) employs late fusion with cross-attention to capture fine-grained semantic relationships, while the adaptive reconciliation module (ARM) applies early fusion with uncertainty embeddings to enhance robustness and adaptability. Experiments on CIRR and FashionIQ show consistent performance gains, improving Recall@1 by up to 2.4% and Recall@50 by 1.04% over existing state-of-the-art methods, empirically demonstrating that FAR Net provides a robust and scalable solution to CIR tasks.

Figures

Figures reproduced from arXiv: 2507.12823 by the authors.

Figure 1
Figure 1. Comparison of existing fusion strategies for CIR. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall training pipeline of the FAR-Net framework. This consists of two main modules: (a) The ESAM extracts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Attention visualization for semantic alignment on (a) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages

  1. [1]

    Composing text and image for image retrieval-an empirical odyssey,

    N. V o et al. , “Composing text and image for image retrieval-an empirical odyssey,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2019, pp. 6439–6448

    work page 2019

  2. [2]

    Content-based retrieval of segmented images,

    T.-S. Chua, S.-K. Lim, and H.-K. Pung, “Content-based retrieval of segmented images,” in Proc. ACM Int. Conf. Multimedia (MM) , 1994, pp. 211–218

    work page 1994

  3. [3]

    Context-aware multi- view summarization network for image-text matching,

    L. Qu, M. Liu, D. Cao, L. Nie, and Q. Tian, “Context-aware multi- view summarization network for image-text matching,” in Proc. ACM Int. Conf. Multimedia (MM) , 2020, pp. 1047–1055

    work page 2020

  4. [4]

    Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things,

    J. Zhang and D. Tao, “Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things,” IEEE Internet Things J. , vol. 8, pp. 7789–7817, 2020

    2020

  5. [5]

    Data roaming and quality assessment for composed image retrieval,

    M. Levy, R. Ben-Ari, N. Darshan, and D. Lischinski, “Data roaming and quality assessment for composed image retrieval,” in Proc. Ann. AAAI Conf. Artif. Intell. (AAAI) , vol. 38, 2024, pp. 2991–2999

    work page 2024

  6. [6]

    CompoDiff: Versatile composed image retrieval with latent diffusion,

    G. Gu et al. , “CompoDiff: Versatile composed image retrieval with latent diffusion,” Trans. Mach. Learn. Res. , 2024

    2024

  7. [7]

    Comprehensive linguistic-visual composition network for image retrieval,

    H. Wen, X. Song, X. Yang, Y . Zhan, and L. Nie, “Comprehensive linguistic-visual composition network for image retrieval,” in Proc. 44th Int. ACM SIGIR Conf. Res. Dev. Inf. Retr . (SIGIR) , 2021, pp. 1369–1378

    work page 2021

  8. [8]

    Simple but effective raw-data level multimodal fusion for composed image retrieval,

    H. Wen et al., “Simple but effective raw-data level multimodal fusion for composed image retrieval,” in Proc. Int. ACM SIGIR Conf. Res. Develop. Inf. Retr . (SIGIR), 2024, pp. 229–239

    work page 2024

Show all 34 references
  1. [9]

    Information captur- ing camera and developmental issues,

    H. Fujisawa, H. Sako, Y . Okada, and S.-W. Lee, “Information captur- ing camera and developmental issues,” in Proc. Int. Conf. Document Anal. Recognit. IEEE, 1999, pp. 205–208

    1999

  2. [10]

    Sentence-level prompts benefit composed image re- trieval,

    X. Xu et al. , “Sentence-level prompts benefit composed image re- trieval,” in Int. Conf. Learn. Represent. (ICLR) , 2024

    2024

  3. [11]

    Dynamic weighted combiner for mixed-modal image retrieval,

    F. Huang, L. Zhang, X. Fu, and S. Song, “Dynamic weighted combiner for mixed-modal image retrieval,” in Proc. Ann. AAAI Conf. Artif. Intell. (AAAI) , vol. 38, 2024, pp. 2303–2311

    2024

  4. [12]

    MMFT- BERT: Multimodal fusion transformer with bert encodings for visual question answering,

    A. U. Khan, A. Mazaheri, N. D. V . Lobo, and M. Shah, “MMFT- BERT: Multimodal fusion transformer with bert encodings for visual question answering,” in Proc. Conf. Empir . Methods Nat. Lang. Process. (EMNLP), 2020, pp. 4648–4660

    2020

  5. [13]

    Multimodal inverse cloze task for knowledge-based visual question answering,

    P. Lerner, O. Ferret, and C. Guinaudeau, “Multimodal inverse cloze task for knowledge-based visual question answering,” in Proc. Eur . Conf. Inf. Retr . (ECIR), 2023, pp. 569–587

    2023

  6. [14]

    Motion influence map for unusual human activity detection and localization in crowded scenes,

    D.-G. Lee, H.-I. Suk, S.-K. Park, and S.-W. Lee, “Motion influence map for unusual human activity detection and localization in crowded scenes,” IEEE Trans. Circuits Syst. Video Technol. , vol. 25, no. 10, pp. 1612–1623, 2015

    2015

  7. [15]

    A new recurrent neural-network archi- tecture for visual pattern recognition,

    S.-W. Lee and H.-H. Song, “A new recurrent neural-network archi- tecture for visual pattern recognition,” IEEE Trans. Neural Networks , vol. 8, no. 2, pp. 331–340, 1997

    1997

  8. [16]

    BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” in Int. Conf. Mach. Learn. (ICML) , 2023, pp. 19 730–19 742

    2023

  9. [17]

    FashionIQ: A new dataset towards retrieving images by natural language feedback,

    H. Wu et al. , “FashionIQ: A new dataset towards retrieving images by natural language feedback,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 11 307–11 317

    2021

  10. [18]

    Image retrieval on real-life images with pre-trained vision-and-language models,

    Z. Liu, C. Rodriguez-Opazo, D. Teney, and S. Gould, “Image retrieval on real-life images with pre-trained vision-and-language models,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 2125–2134

    2021

  11. [19]

    Learning transferable visual models from natural language supervision,

    A. Radford et al. , “Learning transferable visual models from natural language supervision,” in Int. Conf. Mach. Learn. (ICML) , 2021, pp. 8748–8763

    2021

  12. [20]

    FashionVLP: Vision language transformer for fashion retrieval with feedback,

    S. Goenka et al. , “FashionVLP: Vision language transformer for fashion retrieval with feedback,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 14 105–14 115

    2022

  13. [21]

    COSMO: Content-style modulation for image retrieval with text feedback,

    S. Lee, D. Kim, and B. Han, “COSMO: Content-style modulation for image retrieval with text feedback,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 802–812

    2021

  14. [22]

    PAIR: Complementarity-guided disentanglement for composed image retrieval,

    Z. Fu et al. , “PAIR: Complementarity-guided disentanglement for composed image retrieval,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP) , 2025, pp. 1–5

    2025

  15. [23]

    Target-guided composed image retrieval,

    H. Wen, X. Zhang, X. Song, Y . Wei, and L. Nie, “Target-guided composed image retrieval,” in Proc. ACM Int. Conf. Multimedia (MM) , 2023, pp. 915–923

    2023

  16. [24]

    Syncmask: Synchronized attentional masking for fashion-centric vision-language pretraining,

    C. H. Song, T. Hwang, J. Yoon, S. Choi, and Y . H. Gu, “Syncmask: Synchronized attentional masking for fashion-centric vision-language pretraining,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 13 948–13 957

    2024

  17. [25]

    MEDIAN: Adaptive intermediate-grained aggrega- tion network for composed image retrieval,

    Q. Huang et al. , “MEDIAN: Adaptive intermediate-grained aggrega- tion network for composed image retrieval,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP) , 2025, pp. 1–5

    2025

  18. [26]

    Composed image retrieval with text feedback via multi-grained uncertainty regulariza- tion,

    Y . Chen, Z. Zheng, W. Ji, L. Qu, and T.-S. Chua, “Composed image retrieval with text feedback via multi-grained uncertainty regulariza- tion,” in Int. Conf. Learn. Represent. (ICLR) , 2024

    2024

  19. [27]

    Decomposing semantic shifts for composed image retrieval,

    X. Yang, D. Liu, H. Zhang, Y . Luo, C. Wang, and J. Zhang, “Decomposing semantic shifts for composed image retrieval,” in Proc. Ann. AAAI Conf. Artif. Intell. (AAAI) , vol. 38, 2024, pp. 6576–6584

    2024

  20. [28]

    CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning,

    H. Luo et al. , “CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning,” Neurocomputing, vol. 508, pp. 293–304, 2022

    2022

  21. [29]

    ARTEMIS: Attention-based retrieval with text-explicit matching and implicit sim- ilarity,

    G. Delmas, R. S. de Rezende, G. Csurka, and D. Larlus, “ARTEMIS: Attention-based retrieval with text-explicit matching and implicit sim- ilarity,” arXiv preprint arXiv:2203.08101 , 2022

    2022 arXiv

  22. [30]

    An image is worth 16x16 words: transformers for image recognition at scale,

    A. Dosovitskiy et al. , “An image is worth 16x16 words: transformers for image recognition at scale,” in Int. Conf. Learn. Represent. (ICLR) , 2020

    2020

  23. [31]

    Integrated segmentation and recognition of handwritten numerals with cascade neural network,

    S.-W. Lee and S.-Y . Kim, “Integrated segmentation and recognition of handwritten numerals with cascade neural network,” IEEE Trans. Syst. Man Cybern. , vol. 29, no. 2, pp. 285–290, 1999

    1999

  24. [32]

    Accurate object contour tracking based on boundary edge selection,

    M.-C. Roh, T.-Y . Kim, J. Park, and S.-W. Lee, “Accurate object contour tracking based on boundary edge selection,” Pattern Recognit., vol. 40, no. 3, pp. 931–943, 2007

    2007

  25. [33]

    BERT: Pre- training of deep bidirectional transformers for language understand- ing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language understand- ing,” in Proc. Conf. North Amer . Chapter Assoc. Comput. Linguistics: Human Lang. Technol. (NAACL-HLT) , 2019, pp. 4171–4186

    2019

  26. [34]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

    2017 arXiv

Pith tools

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