Pith. sign in

REVIEW 3 major objections 5 minor 86 references

Three Things to Know about Deep Metric Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a training recipe combining a differentiable surrogate of recall@k, a similarity-space mixup that virtually enlarges the batch, and initialization from large-scale pretrained models nearly solves standard deep metric…

desk verdict The recipe nearly solves DML benchmarks, but the loss-vs-Smooth-AP comparison is confounded by batch size, so the loss-superiority claim is unproven. read the letter →

arxiv 2412.12432 v1 pith:476XQ7SL submitted 2024-12-17 cs.CV cs.AI

Yash Patel , Giorgos Tolias , Jiri Matas This is my paper · ORCID
classification cs.CVcs.AI
keywords deepmetriclearningrecall@ksurrogatesimilaritymixuplargebatchtrainingmodelinitializationimageretrievalvisiontransformer
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

Supervised deep metric learning trains an embedding so that nearest-neighbor retrieval works on classes unseen during training. This paper claims that three choices together let large models nearly solve standard benchmarks: a differentiable surrogate of the retrieval metric recall@k trained on very large batches, a mixup that operates on pairwise similarities to virtually enlarge the batch, and initialization from large-scale pretrained vision models. With those ingredients the authors report recall@1 of 90.0% on iNaturalist, 90.8% on Stanford Online Products, and 97.2% on Cars196, with higher recall@k values approaching ceiling on all three. A reader should care because the recipe is simple and, if accurate, turns these long-standing benchmarks into saturated targets that no longer separate methods.

What carries the argument

The machinery is the RS@k loss, a sigmoid-smoothed version of recall@k. In the paper's formulation, the counting of positives inside the top $k$ and the ranking of each positive are both softened by sigmoids, one with temperature $\tau_2 = 0.01$ on similarity differences and one with $\tau_1 = 1$ on rank position, so gradients flow to similarity changes that actually improve the shortlist. Two supporting mechanisms make it work: a multistage backpropagation borrowed from [25] that allows a batch of 4,000 images within GPU memory, and SiMix, which generates virtual examples by linearly mixing the scalar similarities of original same-class pairs, effectively enlarging the batch without creating mixed embeddings. The loss is then applied to each original and virtual image as query, using a set of $k$ values such as $\{1,2,4,8,16\}$.

What would settle it

Train the proposed RS@k and Smooth-AP with identical batch sizes on iNaturalist, SOP, and Cars196 using the same pretrained ViT initialization, and compare recall@1. If the RS@k advantage over Smooth-AP largely disappears when batches are equalized, the central claim that the new loss drives the near-solved results is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the non-differentiable recall@k metric can be optimized directly through a hand-designed smooth surrogate, provided the loss is evaluated on a batch almost as large as the training set. The paper expresses recall@k by replacing its Heaviside step functions with two sigmoids controlled by separate temperatures, and shows that the resulting RS@k loss tracks the true metric closely. Training with batch sizes of 4,000 is made feasible by a multistage backpropagation that recomputes embeddings one at a time, and the SiMix variant further expands the effective batch by mixing scalar similarities of same-class pairs without ever constructing mixed embeddings. Combined with initialization from large-scale pretrained models, this pipeline gives top results on iNaturalist, Stanford Online Products, Cars196, and VehicleID and improves mean Average Precision on instance-level retrieval over an Average Precision loss baseline. The authors conclude that the tested benchmarks are nearly solved under these settings.

Load-bearing premise

The load-bearing premise is that the new loss itself is what produces the performance gains; in most experiments, however, the new loss is trained with a much larger batch than the Smooth-AP baseline, so the improvement could come from batch size rather than from the loss.

Editorial extensions

If this is right

  • Using RS@k instead of Smooth-AP improves recall@1 consistently across ResNet and ViT backbones and embedding sizes in the paper's experiments.
  • SiMix gives its largest gains when the physical batch is small, as on Cars196, so it is a cheap way to recover large-batch behavior under hardware limits.
  • Initialization from CLIP, DINOv2, DiHT, or SWAG generally beats ImageNet-21k initialization, and better pretraining transfers to better retrieval after fine-tuning.
  • At the strongest settings, iNaturalist reaches 97.6% recall@16, SOP reaches 97.7% recall@10, and Cars196 reaches 99.3% recall@8, which the paper takes as nearly solved benchmarks.
  • The loss is robust to the choice of $k$ set and rank temperature, with $\tau_1 = 1.0$ the best setting.

Reading between the lines

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

  • Inference: because most experiments run RS@k at batch size 4,000 while Smooth-AP uses 384, the reported advantage may be partly a batch-size effect; the paper only isolates this on Cars196, so a batch-matched comparison on iNaturalist and SOP would settle attribution.
  • Inference: SiMix's operation purely on scalar similarities could transfer to other listwise or ranking losses, including non-DML retrieval and recommendation, wherever pairwise similarity matrices are available.
  • Inference: if the benchmarks are genuinely near-solved, future work should shift to harder evaluations such as open-vocabulary or long-tail retrieval, where near-ceiling scores cannot signal progress.
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

3 major / 5 minor

Summary. The paper proposes three components for supervised deep metric learning: a differentiable surrogate of recall@k (RS@k), a similarity-mixup virtual batch expansion (SiMix), and the use of large-scale pre-trained initializations. The RS@k loss replaces the Heaviside functions in the definition of recall with two sigmoids having different temperatures, and is trained on very large batches enabled by a two-stage backpropagation trick. SiMix constructs virtual examples by convexly combining embeddings of same-class pairs and computes their similarities algebraically, avoiding explicit mixed embeddings. The paper reports very high recall@1 numbers on iNaturalist, Stanford Online Products, and Cars196 with ViT backbones initialized from CLIP, DiHT, DINOv2, SWAG, and ImageNet-21k, and argues that these benchmarks are nearly solved. The authors also adopt a train/validation split for hyperparameter tuning, which they identify as a fairer protocol than common practice.

Significance. If the central comparison to Smooth-AP were clean, this would be a valuable empirical contribution: the paper ships a systematic study of three factors (loss, batch expansion, initialization) with a careful validation protocol, and the very large batch implementation with the SiMix trick is likely reusable beyond this specific loss. The breadth of backbones and initializations is a useful reference for the community. However, the headline claim that RS@k consistently surpasses Smooth-AP is weakened by a batch-size confound in the main comparisons, and the paper's internal definition of recall@k is inconsistent between the method section and the experiments section. These issues prevent the current version from fully supporting the paper's central claims.

major comments (3)
  1. [Section 4.2/4.3, Table 5, Fig. 9] The comparison between RS@k and Smooth-AP (SAP) is confounded by batch size. Section 4.2 states that SAP uses batch size 384 (the original value), while RS@k uses 4,000 on iNaturalist and SOP, and 392 on Cars196. Figure 9 shows on Cars196 that increasing batch size alone improves r@1 for both SAP and RS@k by several points, with the curves becoming close at the largest batch. Since no SAP-at-4,000 results are reported on iNaturalist, SOP, or VehicleID, the claimed advantage of RS@k over SAP (e.g., iNaturalist R50-512: 68.7 vs 71.2; ViT-B/16 IN-21k: 79.1 vs 83.9) cannot be attributed to the loss rather than the larger batch. Moreover, Table 5 itself contains entries where SAP matches or beats RS@k without SiMix (e.g., CLIP-400M ViT-B/32 Cars196 r@1: SAP 92.6 vs RS@k 91.8; IN-21k ViT-L/16 Cars196: SAP 91.3 vs RS@k 90.4), which contradicts the text's statement that RS@k is 'consistently surpassing' SAP. A matching-batch comparison on the other datasets is required to establish the loss's benefit.
  2. [Section 3.1 Eq. (1) vs Section 4.1] The definition of recall@k is inconsistent between the method and experiments. Equation (1) defines recall@k as the fraction of positive examples retrieved within the top k ranked list, while Section 4.1 defines the evaluation metric r@k as 1 if at least one positive image appears in the top k list, averaged across queries. These are different quantities; the former is a per-query recall of all positives, the latter is a binarized hit-at-k. The proposed surrogate loss in Eq. (5) is derived from Eq. (1), so it optimizes a different objective than the metric reported in the experiments. The paper should clarify which definition is used, how the surrogate relates to the reported metric, and whether the discrepancy affects the conclusions.
  3. [Section 3.3] The SiMix construction treats convex combinations of two same-class embeddings as valid positive virtual examples without re-normalization. Because the original embeddings are L2-normalized, the virtual embeddings in Eq. (8) generally have norm less than 1, so they live on a different norm sphere than the original examples. The paper does not justify this assumption or ablate it against re-normalizing the mixed embeddings (or against using mixed embeddings only in the database, not as queries). Since the efficiency claim of SiMix relies on this design choice, an analysis or comparison of these variants would strengthen the contribution.
minor comments (5)
  1. [Abstract/Conclusions] The text says 'we archive 97.6% Recall@16' in the introduction and 'faciliated' in the conclusions; these appear to be typos for 'achieve' and 'facilitated'.
  2. [Figure 9] The x-axis labels '24 25 26 27 28 29' are unclear; if they denote powers of two (16, 32, 64, ...), the axis should be labeled accordingly (e.g., 2^4, 2^5) for readability.
  3. [Section 3.3 and throughout] The method is called 'SiMix' in some places and 'SimMix' in others (e.g., Algorithm 2, Figures 5 and 8); the spelling should be unified.
  4. [Abstract] The abstract claims the loss is computed on batches 'nearly equivalent to the entire training set', but the maximum reported batch size is 4,000, which is far smaller than the training sets of iNaturalist (325k) or GLDv1 (1M). This phrasing is an overstatement.
  5. [Table 5 and Section 4.3] The text states that DINOv2 is the best performing on iNaturalist 'by a large margin', but Table 5 shows the gap to CLIP-2B ViT-L/14 is about 7 points at r@1 (90.0 vs 82.8); while substantial, the claim could be more precise about which comparison is meant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper makes empirical benchmark claims with a fully specified surrogate loss, evaluated on held-out test classes.

full rationale

The paper makes empirical claims about a training recipe (RS@k loss + SiMix + pretrained initialization). The central results are recall@1 scores computed with the actual evaluation metric on held-out test classes, not quantities derived from the loss by construction. The proposed loss in Eq. (5) is a smooth approximation of the recall@k formula in Eq. (3), obtained by replacing Heaviside functions with sigmoids; this is the intended surrogate relationship, not a circular prediction, because the reported test scores are measured after training and are not fed back into the loss. Hyperparameters such as tau1, tau2, and the set K are stated as fixed or studied on a validation split; the paper explicitly introduces a 'fair validation protocol' to avoid tuning on the test set. The self-citation to prior work [24] is acknowledged ('extension of our prior work [24]') and the method is fully specified in the paper itself, so no load-bearing claim is outsourced to an unverifiable self-citation. The comparison with Smooth-AP is conducted with unequal batch sizes (SAP at 384 vs RS@k at 4,000 on most datasets, with Fig. 9 showing batch-size sensitivity on Cars196); this is a legitimate experimental confound for the loss-superiority claim, but it is not a construction-level circularity. No equation in the paper reduces to its own input, no fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported from the authors' prior work. Therefore the circularity score is 0.

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

The central empirical claims rest on standard machine-learning assumptions: sigmoid surrogates approximate the step function, large batches approximate full-database recall, and pre-trained features transfer to retrieval. The only paper-specific design axiom is SiMix's use of non-normalized convex combinations of same-class embeddings as valid positives. No new physical entities or theoretical constants are introduced.

free parameters (5)
  • tau1 (rank sigmoid temperature) = 1.0
    Temperature for the sigmoid counting positives within top k in Eq. (5); tuned on Cars196 in Fig. 8 and fixed across datasets.
  • tau2 (similarity difference sigmoid temperature) = 0.01
    Temperature for the sigmoid approximating the rank in Eq. (5); set equal to Smooth-AP [49] and not tuned.
  • k set for multi-k loss = {1,2,4,8,16} without SiMix; {1,2,4,8,12,16,20,24,28,32} with SiMix
    Set of recall cutoffs K in Eq. (7); chosen per configuration and evaluated in Table 3.
  • learning rate, decay steps, decay size, epochs = per-dataset grid search results
    Selected by randomized grid search on a validation split derived from the training set (Section 4.2).
  • samples per class = 4
    Class-balanced sampling with 4 images per class and M classes in a batch; standard in prior DML work.
assumptions (5)
  • standard math The logistic sigmoid with temperature tau is a differentiable approximation of the Heaviside step function, and its gradient provides useful learning signal (Eq. 4).
    Invoked in Section 3.2 to define the surrogate recall in Eq. (5).
  • domain assumption Computing the surrogate loss on a large mini-batch approximates computing it on the full training database.
    Section 3.4 uses this to justify huge batch sizes and the multi-stage backprop algorithm.
  • domain assumption Optimizing a smooth surrogate of recall@k by gradient descent transfers to test-time retrieval performance.
    Central training hypothesis evaluated throughout Section 4; not proven but empirically supported.
  • domain assumption For every query, relevance is defined by class labels and recall@k as in Eq. (1) is the objective of interest, while the published benchmarks use the 'at least one positive in top k' variant.
    Section 3.1 defines the task; Section 4.1 notes the benchmark definition differs from Eq. (1).
  • ad hoc to paper Convex combinations of two same-class embeddings, without re-normalization, form valid positive virtual examples for training.
    SiMix design in Section 3.3, Eqs. (8)-(10); the paper chooses not to re-normalize mixed embeddings for efficiency, which is different from prior mixup practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Three Things to Know about Deep Metric Learning." pith.science (2026). https://pith.science/paper/476XQ7SL

@misc{pith2026241212432,
  author       = {Pith},
  title        = {Pith review of: Three Things to Know about Deep Metric Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/476XQ7SL}},
  note         = {Machine review of arXiv:2412.12432}
}
read the original abstract

This paper addresses supervised deep metric learning for open-set image retrieval, focusing on three key aspects: the loss function, mixup regularization, and model initialization. In deep metric learning, optimizing the retrieval evaluation metric, recall@k, via gradient descent is desirable but challenging due to its non-differentiable nature. To overcome this, we propose a differentiable surrogate loss that is computed on large batches, nearly equivalent to the entire training set. This computationally intensive process is made feasible through an implementation that bypasses the GPU memory limitations. Additionally, we introduce an efficient mixup regularization technique that operates on pairwise scalar similarities, effectively increasing the batch size even further. The training process is further enhanced by initializing the vision encoder using foundational models, which are pre-trained on large-scale datasets. Through a systematic study of these components, we demonstrate that their synergy enables large models to nearly solve popular benchmarks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

86 extracted references · 73 canonical work pages

  1. [24]

    & Matas, J.Recall@ k surro- gate loss with large batches and similarity mixup (2022)

    Patel, Y ., Tolias, G. & Matas, J.Recall@ k surro- gate loss with large batches and similarity mixup (2022)

    work page 2022

  2. [1]

    & Huang, T

    Yu, J., Jiang, Y ., Wang, Z., Cao, Z. & Huang, T. Unitbox: An advanced object detection network (2016)

    2016

  3. [2]

    Rezatofighi, H. et al. Generalized intersection over union: A metric and a loss for bounding box regression (2019)

    2019

  4. [3]

    Nagendar, G., Singh, D., Balasubramanian, V . N. & Jawahar, C. Neuro-iou: Learning a surrogate loss for semantic segmentation. (2018)

    2018

  5. [4]

    & Van Gool, L

    Mentzer, F., Agustsson, E., Tschannen, M., Tim- ofte, R. & Van Gool, L. Conditional probability models for deep image compression (2018)

    2018

  6. [5]

    Ball ´e, J., Minnen, D., Singh, S., Hwang, S. J. & Johnston, N. Variational image compression with a scale hyperprior (2018)

    2018

  7. [6]

    & Manmatha, R

    Patel, Y ., Appalaraju, S. & Manmatha, R. Saliency driven perceptual image compression (2021). 12 Arch. Loss Train-set Mean RO RO+R1M RPar RP+R1M all R1M med hard med hard med hard med hard GeM∗ AP [48] Landmarks-clean [83][84] [25]/ [73] 49.7 36.7 67.1 42.3 47.8 22.5 80.3 60.9 51.9 24.6 GeM∗ AP [48] GLDv1 [72] [25]/github - - 66.3 42.5 - - 80.2 60.8 - - ...

    2021

  8. [7]

    & LeCun, Y .Dimension- ality reduction by learning an invariant mapping (2006)

    Hadsell, R., Chopra, S. & LeCun, Y .Dimension- ality reduction by learning an invariant mapping (2006)

    work page 2006

Show all 86 references
  1. [8]

    & Philbin, J

    Schroff, F., Kalenichenko, D. & Philbin, J. Facenet: A unified embedding for face recogni- tion and clustering (2015)

    2015

  2. [9]

    Wu, C.-Y ., Manmatha, R., Smola, A. J. & Kra- henbuhl, P. Sampling matters in deep embedding learning (2017)

    2017

  3. [10]

    Bahdanau, D. et al. An actor-critic algorithm for sequence prediction (2017)

    2017

  4. [11]

    & Matas, J

    Patel, Y ., Hoda ˇn, T. & Matas, J. Learning surrogates via deep embedding (2020)

    2020

  5. [12]

    & Cord, M

    Engilberge, M., Chevallier, L., P ´erez, P. & Cord, M. Sodeep: a sorting deep net to learn ranking loss surrogates (2019)

    2019

  6. [13]

    Van Horn, G. et al. The inaturalist species classification and detection dataset (2018)

    2018

  7. [14]

    & Bober, M

    Ong, E.-J., Husain, S. & Bober, M. Siamese net- work of deep fisher-vector descriptors for image retrieval (2017)

    2017

  8. [15]

    & Fei-Fei, L

    Krause, J., Stark, M., Deng, J. & Fei-Fei, L. 3d object representations for fine-grained catego- rization (2013)

    2013

  9. [16]

    & Fei- Fei, L

    Dong, W., Socher, R., Li-Jia, L., Li, K. & Fei- Fei, L. ImageNet: A large-scale hierarchical image database (2009)

    2009

  10. [17]

    Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale (2021)

    2021

  11. [18]

    Singh, M. et al. Revisiting weakly supervised pre-training of visual perception models (2022)

    2022

  12. [19]

    Radenovic, F. et al. Filtering, distillation, and hard negatives for vision-language pre-training (2023)

    2023

  13. [20]

    Radford, A. et al. Learning transferable visual models from natural language supervision (2021)

    2021

  14. [21]

    Oquab, M. et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  15. [22]

    Caron, M. et al. Emerging properties in self- supervised vision transformers (2021)

    2021

  16. [23]

    Caron, M. et al. Unsupervised learning of visual features by contrasting cluster assign- ments. NeurIPS (2020)

    2020

  17. [25]

    Revaud, J., Almaz ´an, J., Rezende, R. S. & Souza, C. R. d. Learning with average precision: Train- ing image retrieval with a listwise loss (2019). 13 Method Arch.dim iNaturalist [13] SOP [14] VehicleID [70] Cars196 [15] Small Medium Large 1 4 16 32 100 101 102 103 1 5 1 5 1 ...

    2019

  18. [26]

    Zhang, H., Cisse, M., Dauphin, Y . N. & Lopez- Paz, D. mixup: Beyond empirical risk minimiza- tion. arXiv preprint arXiv:1710.09412 (2017)

    2017 arXiv

  19. [27]

    & Wu, H.-Y

    Zhai, A. & Wu, H.-Y . Classification is a strong baseline for deep metric learning. arXiv preprint arXiv:1811.12649 (2018)

    2018 arXiv

  20. [28]

    Liu, W. et al. Sphereface: Deep hypersphere embedding for face recognition (2017)

    2017

  21. [29]

    Wang, H. et al. Cosface: Large margin cosine loss for deep face recognition (2018)

    2018

  22. [30]

    & Zafeiriou, S

    Deng, J., Guo, J., Xue, N. & Zafeiriou, S. Arc- face: Additive angular margin loss for deep face recognition (2019)

    2019

  23. [31]

    K., Ioffe, S

    Movshovitz-Attias, Y ., Toshev, A., Leung, T. K., Ioffe, S. & Singh, S. No fuss distance metric learning using proxies (2017)

    2017

  24. [32]

    W., DeVries, T

    Teh, E. W., DeVries, T. & Taylor, G. W. Prox- ynca++: Revisiting and revitalizing proxy neigh- borhood component analysis (2020)

    2020

  25. [33]

    Qian, Q. et al. Softtriple loss: Deep metric learning without triplet sampling (2019)

    2019

  26. [34]

    & Kwak, S

    Kim, S., Kim, D., Cho, M. & Kwak, S. Proxy anchor loss for deep metric learning (2020)

    2020

  27. [35]

    & Leal-Taix´e, L

    Elezi, I., Vascon, S., Torcinovich, A., Pelillo, M. & Leal-Taix´e, L. The group loss for deep metric learning (2020)

    2020

  28. [36]

    & Leal-Taix ´e, L

    Seidenschwarz, J., Elezi, I. & Leal-Taix ´e, L. Learning intra-batch connections for deep met- ric learning (2021)

    2021

  29. [37]

    & Ommer, B

    Kotovenko, D., Ma, P., Milbich, T. & Ommer, B. Cross-image-attention for conditional embed- dings in deep metric learning (2023). 14 Init. Arch.dim Method iNaturalist [13] SOP [14] Cars196 [15] 1 4 16 32 10 0 101 102 103 1 2 4 8 IN-21K ViT-B/32512 init 46.9 64 .1 78 .2 83 .7 5...

    2023

  30. [38]

    Boudiaf, M. et al. Metric learning: cross-entropy vs. pairwise losses (2020)

    2020

  31. [39]

    & Lim, S.-N

    Musgrave, K., Belongie, S. & Lim, S.-N. A metric learning reality check (2020)

    2020

  32. [40]

    Roth, K. et al. Revisiting training strategies and generalization performance in deep metric learning (2020)

    2020

  33. [41]

    O., Xiang, Y ., Jegelka, S

    Song, H. O., Xiang, Y ., Jegelka, S. & Savarese, S. Deep metric learning via lifted structured feature embedding (2015)

    2015

  34. [42]

    Improved deep metric learning with multi-class n-pair loss objective (2016)

    Sohn, K. Improved deep metric learning with multi-class n-pair loss objective (2016)

    2016

  35. [43]

    & Lin, Y

    Wang, J., Zhou, F., Wen, S., Liu, X. & Lin, Y . Deep metric learning with angular loss (2017)

    2017

  36. [44]

    Wen, Y . et al. Pairwise similarity learning is simple (2023)

    2023

  37. [45]

    & Lempitsky, V

    Ustinova, E. & Lempitsky, V . Learning deep embeddings with histogram loss (2016)

    2016

  38. [46]

    Kemertas, M., Pishdad, L., Derpanis, K. G. & Fazly, A. Rankmi: A mutual information maxi- mizing ranking loss (2020)

    2020

  39. [47]

    Rol ´ınek, M. et al. Optimizing rank-based metrics with blackbox differentiation (2020). 15

    2020

  40. [48]

    & Sclaroff, S

    He, K., Lu, Y . & Sclaroff, S. Local descriptors optimized for average precision (2018)

    2018

  41. [49]

    & Zisser- man, A

    Brown, A., Xie, W., Kalogeiton, V . & Zisser- man, A. Smooth-ap: Smoothing the path towards large-scale image retrieval (2020)

    2020

  42. [50]

    & Bitot, X

    Ramzi, E., Thome, N., Rambour, C., Audebert, N. & Bitot, X. Robust and decomposable aver- age precision for image retrieval(2021)

    2021

  43. [51]

    & Bitot, X

    Ramzi, E., Audebert, N., Thome, N., Rambour, C. & Bitot, X. Hierarchical average precision training for pertinent image retrieval (2022)

    2022

  44. [52]

    Verma, V . et al. Manifold mixup: Better repre- sentations by interpolating hidden states (2019)

    2019

  45. [53]

    & Zhou, J

    Duan, Y ., Zheng, W., Lin, X., Lu, J. & Zhou, J. Deep adversarial metric learning (2018)

    2018

  46. [54]

    B., Pion, N., Wein- zaepfel, P

    Kalantidis, Y ., Sariyildiz, M. B., Pion, N., Wein- zaepfel, P. & Larlus, D. Hard negative mixing for contrastive learning. NeurIPS (2020)

    2020

  47. [55]

    & Zhou, J.Hardness- aware deep metric learning (2019)

    Zheng, W., Chen, Z., Lu, J. & Zhou, J.Hardness- aware deep metric learning (2019)

    2019

  48. [56]

    Gu, G. & Ko, B. Symmetrical synthesis for deep metric learning (2020)

    2020

  49. [57]

    & Kim, H.-G

    Gu, G., Ko, B. & Kim, H.-G. Proxy synthesis: Learning with synthetic classes for deep metric learning. AAAI (2021)

    2021

  50. [58]

    Venkataramanan, S. et al. It takes two to tango: Mixup for deep metric learning. arXiv preprint arXiv:2106.04990 (2021)

    2021 arXiv

  51. [59]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition (2016)

    2016

  52. [60]

    & Markov, S

    Kyurkchiev, N. & Markov, S. Sigmoid functions: some approximation and modelling aspects.LAP LAMBERT Academic Publishing, Saarbrucken (2015)

    2015

  53. [61]

    I., Kyurkchiev, N

    Iliev, A. I., Kyurkchiev, N. & Markov, S. On the approximation of the cut and step functions by logistic and gompertz functions. Biomath (2015)

    2015

  54. [62]

    & Markov, S

    Iliev, A., Kyurkchiev, N. & Markov, S. On the approximation of the step function by some sig- moid functions. Mathematics and Computers in Simulation (2017)

    2017

  55. [63]

    & Hinton, G

    Salakhutdinov, R. & Hinton, G. Semantic hash- ing. International Journal of Approximate Rea- soning (2009)

    2009

  56. [64]

    & Mnih, A

    Gu, S., Levine, S., Sutskever, I. & Mnih, A. Muprop: Unbiased backpropagation for stochas- tic neural networks (2016)

    2016

  57. [65]

    J., Mnih, A

    Maddison, C. J., Mnih, A. & Teh, Y . W. The concrete distribution: A continuous relaxation of discrete random variables (2017)

    2017

  58. [66]

    Simo-Serra, E. et al. Discriminative learning of deep convolutional feature point descriptors (2015)

    2015

  59. [67]

    Zhou, J. et al. ibot: Image bert pre- training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)

    2021 arXiv

  60. [68]

    Ilharco, G. et al. Openclip (2021). URL https: //doi.org/10.5281/zenodo.5143773

    2021 doi

  61. [69]

    Cherti, M. et al. Reproducible scaling laws for contrastive language-image learning (2023)

    2023

  62. [70]

    & Huang, T

    Liu, H., Tian, Y ., Wang, Y ., Pang, L. & Huang, T. Deep relative distance learning: Tell the difference between similar vehicles (2016)

    2016

  63. [71]

    & Chum, O

    Radenovi ´c, F., Iscen, A., Tolias, G., Avrithis, Y . & Chum, O. Revisiting oxford and paris: Large- scale image retrieval benchmarking (2018)

    2018

  64. [72]

    & Han, B

    Noh, H., Araujo, A., Sim, J., Weyand, T. & Han, B. Large-scale image retrieval with attentive deep local features (2017)

    2017

  65. [73]

    & Chum, O

    Tolias, G., Jenicek, T. & Chum, O. Learning and aggregating deep local descriptors for instance- level recognition (2020)

    2020

  66. [74]

    & Chum, O

    Radenovi ´c, F., Tolias, G. & Chum, O. Fine- tuning cnn image retrieval with no human anno- tation. PAMI (2019)

    2019

  67. [75]

    L., Kiros, J

    Ba, J. L., Kiros, J. R. & Hinton, G. E. Layer normalization. arXiv preprint arXiv:1607.06450 16 (2016)

    2016 arXiv

  68. [76]

    Pytorch image models

    Wightman, R. Pytorch image models. https: //github.com/rwightman/pytorch-image-models (2019)

    2019

  69. [77]

    Kingma, D. & Ba, J. Adam: A method for stochastic optimization (2015)

    2015

  70. [78]

    & Hutter, F

    Loshchilov, I. & Hutter, F. Decoupled weight decay regularization (2019)

    2019

  71. [79]

    Szegedy, C. et al. Going deeper with convolu- tions (2015)

    2015

  72. [80]

    & Sclaroff, S

    Cakir, F., He, K., Xia, X., Kulis, B. & Sclaroff, S. Deep metric learning to rank (2019)

    2019

  73. [81]

    & Scott, M

    Wang, X., Zhang, H., Huang, W. & Scott, M. R. Cross-batch memory for embedding learning (2020)

    2020

  74. [82]

    & Klein, E

    Jacob, P., Picard, D., Histace, A. & Klein, E. Metric learning with horde: High-order regular- izer for deep embeddings (2019)

    2019

  75. [83]

    & Lem- pitsky, V

    Babenko, A., Slesarev, A., Chigorin, A. & Lem- pitsky, V . Neural codes for image retrieval (2014)

    2014

  76. [84]

    & Larlus, D

    Gordo, A., Almazan, J., Revaud, J. & Larlus, D. End-to-end learning of deep visual representa- tions for image retrieval. IJCV (2017)

    2017

  77. [85]

    & Ommer, B

    Roth, K., Brattoli, B. & Ommer, B. Mic: Mining interclass characteristics for improved metric learning (2019)

    2019

  78. [86]

    & Scott, M

    Wang, X., Han, X., Huang, W., Dong, D. & Scott, M. R. Multi-similarity loss with general pair weighting for deep metric learning (2019). 17

    2019

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.