Pith. sign in

REVIEW 2 major objections 4 minor 31 references

Rethinking Multi-Branch and Cross-Backbone Fusion for Vehicle Re-Identification in the Foundation-Model Era

T0 review · 2 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read At foundation-model scale, multi-branch and cross-backbone fusion stop paying off for vehicle re-identification: a single tuned ConvNeXt backbone matches the strongest multi-branch baseline, and fusion is bounded to +0.11 mAP.

desk verdict A careful, well-evidenced negative result on fusion for vehicle Re-ID at foundation scale, but the tight quantitative bounds rest on single-seed runs and should be treated as provisional. read the letter →

arxiv 2607.22068 v1 pith:SC4U6HDN submitted 2026-07-24 cs.CV cs.LG

classification cs.CVcs.LG
keywords vehiclere-identificationmulti-branchfusioncross-backbonefoundationmodelsDINOv3ConvNeXtre-rankingrepresentationaldiversity
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

The paper argues that the long-standing assumption behind multi-branch vehicle re-identification—that combining complementary feature streams improves retrieval—breaks down when the backbone is a foundation model. A single DINOv3-pretrained ConvNeXt, trained with a tuned recipe and using only visual cues, reaches 88.19 mAP on VeRi-Wild Small, matching the strongest published multi-branch baseline that also uses camera and viewpoint metadata. Concatenating four heads over one shared backbone changes the best single-head result by less than one mAP point while quadrupling the embedding dimension. Even when fusing ConvNeXt with a Vision Transformer under maximally favorable conditions—a frozen anchor, three training recipes, and a LoRA control—the paired bootstrap bounds any fusion gain at +0.11 mAP. The authors conclude that for this setting, spending compute on one strong backbone plus re-ranking is more effective than adding branches or a second backbone.

What carries the argument

The argument rests on a simple identity: for unit-normalized branch embeddings, concatenating and re-normalizing yields a cosine similarity that is the arithmetic mean of the per-branch cosines. Consequently, concatenation can beat its best branch only if the branches disagree in their retrieval orderings (Proposition 1). This turns retrieval-level diagnostics—top-k Jaccard overlap, complementary-correctness, and occlusion-saliency agreement—into direct measures of the headroom available to fusion. To test cross-backbone fusion fairly, the paper uses an asymmetric frozen-anchor scheme: the converged 88.19 mAP ConvNeXt is frozen as a semantic anchor, and only the Vision Transformer and fusion

What would settle it

Train the same ConvNeXt and ViT-L backbones from independent pretraining lineages (one supervised, one self-supervised, not both distilled from the same DINOv3 teacher) and fuse them with the frozen-anchor scheme; a positive net gain over the stronger backbone's solo ceiling beyond the +0.11 mAP paired 95% upper bound would falsify the cross-backbone claim. The paper itself lists this as falsifier (a), along with (c) a ViT recipe that closes the 13–15 mAP gap at this data scale.

Watch

Extended reading notes

Core claim

The paper's central claim is negative: representational diversity, the principle that drove a decade of multi-branch Re-ID designs, no longer yields measurable gains when backbones are foundation-model-scale. The evidence is twofold. First, same-backbone heads collapse into redundancy: four heads on a shared DINOv3-ConvNeXt become more alike as training proceeds (top-10 Jaccard rises from 0.79 to 0.84, occlusion-heatmap agreement reaches 0.96), and the 8192-d concatenated embedding has an effective rank of about 512—four branches produce one representation. Second, cross-backbone heterogeneity is real but unfusible: a frozen ConvNeXt and ViT-L differ genuinely (occlusion correlation 0.30, Ja

Load-bearing premise

The headline quantitative claims—parity within 0.07 mAP, the −0.05 concatenation delta, and the +0.03 fusion delta—come from single-seed training, and the bootstrap intervals capture evaluation uncertainty only, not training-seed variance; a different seed could move these sub-0.1 mAP numbers, though the qualitative negative result is buttressed by three ViT recipes, a LoRA control, and the oracle weight-zero finding.

Editorial extensions

If this is right

  • Under the evaluated setting, multi-branch concatenation adds less than one mAP point over the best single head while costing 4x the embedding dimension, so practitioners should not expect diversity from shared-backbone branches.
  • Cross-backbone CNN–Transformer fusion, even given every structural advantage, is bounded above by +0.11 mAP (paired 95% CI) and costs 2.8x latency and 3x FLOPs, making it dominated by a single tuned backbone plus re-ranking.
  • A tuned single backbone trained purely on visual signals reaches parity with the strongest protocol-verified metadata-dependent multi-branch baseline, suggesting that camera/view annotations are not necessary ingredients at this scale.
  • Protocol discipline matters: using the official cross-camera junk filter changes mAP by 3–4 points; unfiltered numbers can silently misrank methods, so all comparisons should be protocol-matched.
  • The released exact sparse re-ranking implementation removes the memory wall for k-reciprocal re-ranking on large galleries, making a training-free retrieval-stage gain of +4–6 mAP practical at scale.

Reading between the lines

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

  • If this pattern holds beyond the one foundation-model family and two benchmarks tested, the multi-branch design pattern for Re-ID should be re-evaluated: the field's budget should shift from architectural complexity toward recipe tuning, data, and retrieval-stage processing.
  • The finding that fusion appears most beneficial when the second branch is weakest suggests that some published fusion gains may be artifacts of undertrained single-backbone baselines; re-benchmarking prior fusion methods against independently tuned single backbones could change leaderboard conclusions.
  • The convergence-only measurement rule is a methodological recommendation that could transfer to other retrieval tasks: branch-level or fusion claims should be evaluated only at convergence, since undertrained snapshots here inverted every headline conclusion.
  • The exact sparse re-ranking technique is likely to benefit other large-gallery retrieval problems (e.g., person re-identification, landmark retrieval, near-duplicate search) where the classical dense implementation is memory-prohibitive.
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

2 major / 4 minor

Summary. The paper revisits the long-standing assumption that multi-branch and CNN–Transformer feature diversity improves vehicle re-identification, in the era of foundation models. Using the official cross-camera protocol on VeRi-Wild and VeRi-776, the authors show that a single DINOv3-pretrained ConvNeXt-Base with a tuned recipe reaches 88.19/77.47 mAP, matching the strongest protocol-verified metadata-dependent multi-branch baseline MBR-4B-LAI, and rises to 92.38/83.68 with a training-free exact sparse re-ranking. Through branch-level diagnostics (top-k Jaccard, occlusion saliency, complementary correctness, linear CKA), the paper reports that shared-backbone heads become redundant at convergence and that cross-backbone ConvNeXt/ViT fusion, even under an asymmetric frozen-anchor scheme designed to favor fusion, yields no significant gain, with a paired per-query bootstrap upper bound of +0.11 mAP. The paper releases code, checkpoints, and a diagnostic toolkit, and it states explicit falsifiers and scopes its claims to single-seed training and one foundation-model family.

Significance. If the empirical conclusions hold, the paper makes a practically important and field-relevant contribution: at foundation-model scale and with a strong pretraining lineage, a single well-tuned backbone plus retrieval-stage re-ranking is the efficiency frontier, and a decade of multi-branch and dual-backbone fusion designs may not deliver measurable retrieval gains. Methodologically, the paper is unusually careful: the official protocol is verified against the original MBR evaluation code with the measured inflation (+3.04 VeRi-776, +3.94 VeRi-Wild) quantified; convergence-only measurement is adopted after showing that early snapshots reverse all conclusions; paired per-query bootstraps are used to adjudicate small deltas; oracle score-level fusion and a shared-teacher control address obvious confounds; and the explicit release of code, checkpoints, and falsifiers makes the work reproducible and the claims checkable. These strengths are substantial and set a high bar for empirical work in this area.

major comments (2)
  1. [Section IV-B and Section VII] The headline bound 'never exceeds its own single-backbone component beyond a paired 95% upper bound of +0.11 mAP' is a stability claim about the training distribution, but the evidence comes from a single training seed. The paired per-query bootstrap (CI [-0.05, +0.11] around the +0.03 delta) quantifies only evaluation uncertainty conditional on the trained model, as the authors explicitly acknowledge in Section VII. If seed-level mAP variation from random initialization and data sampling is on the order of the unpaired CI (±0.66 on VeRi-Wild Small), the +0.03 observed delta could be negative or substantially positive in another seed. The same single-seed limitation affects the −0.05 concatenation delta and the 0.07 parity margin against MBR-4B-LAI. Because these quantities are load-bearing for the abstract and conclusion, the paper should either (i) run 3–5 seeds for Option A and the as
  2. [Section IV-B] The parity claim with MBR-4B-LAI also rests on a single seed. The margin is 0.07 mAP on Small and 0.06 on Large, while the unpaired query-level bootstrap CI has half-widths of ±0.66 and ±0.51. The authors are careful to claim only parity, not superiority, and this is appropriate. However, the phrase 'matches the strongest protocol-verified metadata-dependent multi-branch baseline' in the abstract could be read as a general property of the method rather than of one training run. If seed variance is comparable to the CI, a different seed could place the method 0.5 mAP below the baseline. The authors should either provide multi-seed evidence for the parity conclusion or explicitly state in the abstract that the parity is for the released single checkpoint. This is not a reason to reject, but it is part of the same single-seed problem that needs to be resolved before the stated claims can be
minor comments (4)
  1. [Section I] The word 'na¨ıve' contains a formatting artifact; please render as 'naive' or 'naïve' consistently.
  2. [Section V-B] Typo: 'One One component does not carry over' should read 'One component does not carry over.'
  3. [Table 4] The phrase 'Diversity vanishes—in reverse—as training proceeds' is confusing; consider rewording to 'Diversity decreases as training proceeds, the opposite of the diversity premise.'
  4. [Section VI] The efficiency numbers (latency, FLOPs, parameters) are clearly stated, but the inference-time parameter count for the dual-backbone system excludes classifier heads; please make explicit whether the same exclusion applies to the single-backbone number for an apples-to-apples comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all central claims are benchmark measurements against external baselines; Proposition 1 is an elementary identity used for diagnostics, and the single-seed caveat is a validity limitation, not a circular step.

full rationale

No material circularity found. The paper's headline claims are empirical measurements compared against external published baselines (MBR-4B-LAI and other prior work), not derivations from fitted inputs. Proposition 1 is an elementary identity about mean cosine similarities and ranking metrics; it is used only to justify retrieval-level diagnostics and is not used to manufacture the null fusion results. The frozen-anchor cross-backbone fusion does contain the strong ConvNeXt weights, but the paper explicitly does not claim a by-construction floor: 'We therefore do not claim a by-construction floor of 88.19', and the fused output is a different 1024-d readout that must be learned; the +0.11 mAP upper bound is a paired bootstrap over observed fused versus single-backbone scores. The oracle score-level fusion is an empirical upper-bound search that happens to select weight zero; it is a measurement, not an assumption. Citations such as [14] and [10] provide external mechanistic explanations for the observed collapse and do not constitute a load-bearing self-citation chain; no uniqueness theorem or ansatz is imported from the authors' prior work. The single-seed limitation stated in Section VII ('We did not run multi-seed retraining, so every sub-0.1 mAP comparison ... is a single-seed observation') is a statistical-validity caveat about training-seed variance, not a circularity; the qualitative redundancy and fusion-null pattern is independently corroborated by multiple recipes, a LoRA control, and the oracle result. The derivation chain is therefore self-contained with respect to circularity, even though the numerical tightness of some claims is limited by single-seed evidence.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The paper's claims rest on training-protocol choices and statistical assumptions rather than on fitted constants: no parameter is adjusted to make the negative result appear, and the negative findings held across dozens of configurations. The main unmeasured quantity is seed-level variance: all sub-0.1 mAP deltas are single-seed observations and the bootstrap intervals bound query-level, not training-level, uncertainty. Loss weights and some schedule details are absent from the text, which is the clearest gap between the stated recipe and a fully reproducible one. No invented entities (forces, mediators, dimensions) are postulated; the frozen-anchor scheme is a training procedure, not a new entity.

free parameters (7)
  • Staged LR decay (0.1x factor at epoch 40) = 0.1 factor at epoch 40
    Hand-set recipe choice attributed +2.21 mAP (§IV-A, §V-B); central to the single-backbone parity claim.
  • Frozen warm-up length / unfreezing schedule = warm-up ≈5 epochs; unfreeze jump +16.8 mAP
    Recipe component with a step-like effect on the learning curve (§IV-A).
  • Combined loss weights (CE, Circle, Center, Triplet)
    Option A is trained with all four losses (§IV-A) but relative weights are not stated — a reproducibility gap.
  • Fusion ViT learning-rate variants = 2.5e-4 base; ×0.1, ×0.25, ×0.45; layer-wise peak 3e-4, 0.98/layer
    Grid of three full-fine-tuning recipes used to establish the ~73 mAP ViT ceiling (Table 6).
  • LoRA configuration for ViT control = rank 8, alpha 16, blocks 20–23
    PEFT control reaching 75.19 mAP (§V-C); given in the text.
  • Re-ranking hyperparameters = k1=20, k2=6, lambda=0.3
    Inherited unchanged from [19]; contributes +4.19/+6.21 mAP; not refit here, listed for completeness.
  • 4-head topology choices (pure-ConvNeXt heads; full-depth trunk)
    Head type and trunk depth are recipe choices; replacing ConvNeXt heads with ResNet-Bottleneck heads costs +6.4 mAP, making this choice load-bearing for the parity claim (§V-B).
assumptions (6)
  • domain assumption Official cross-camera junk-filter protocol is correct, and the authors' reproduction matches the original MBR evaluation code's behavior exactly
    Section III; all mAP figures inherit this. If the published MBR numbers used a different protocol, the parity claim is void; the paper measures unfiltered inflation at +3.04/+3.94 mAP.
  • domain assumption Query-level paired bootstrap captures the relevant uncertainty for the sub-0.1 mAP comparisons
    Sections IV-B and V-C; the ±0.66 vs ±0.14 CI logic is valid only if queries are the right exchangeable unit and seed variance is negligible.
  • domain assumption Full fine-tuning of self-supervised ViTs is fragile (catastrophic forgetting [10], [29]); LoRA partially mitigates it
    Section V-C; used to interpret the ~73 full-fine-tuning ceiling as partly an artifact. The '13–15 mAP gap' conclusion depends on the adaptation regime being representative.
  • domain assumption Linear CKA on 2,500 frozen, global-average-pooled images measures heterogeneity relevant to fusion headroom
    Section V-D; the authors themselves call CKA 'a proxy' and note cross-lineage cells mix preprocessing regimes.
  • standard math Proposition 1: concatenation of unit-normalized branches equals the mean cosine; identical per-branch rankings imply no concatenation gain
    Section III; elementary but correctly stated; motivates all retrieval-level diagnostics.
  • domain assumption Findings transfer from VeRi-Wild to VeRi-776 and to mid-sized benchmarks at 256px within the DINOv3 family
    Section VII; explicitly scoped, and the authors' own falsifiers (a)–(e) describe where transfer could fail.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Multi-Branch and Cross-Backbone Fusion for Vehicle Re-Identification in the Foundation-Model Era." pith.science (2026). https://pith.science/paper/SC4U6HDN

@misc{pith2026260722068,
  author       = {Pith},
  title        = {Pith review of: Rethinking Multi-Branch and Cross-Backbone Fusion for Vehicle Re-Identification in the Foundation-Model Era},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SC4U6HDN}},
  note         = {Machine review of arXiv:2607.22068}
}
read the original abstract

Multi-branch architectures and CNN-Transformer fusion have long been regarded as effective ways to improve vehicle re-identification (Re-ID) by combining complementary representations. In this work, we revisit this assumption in the foundation-model era through a comprehensive empirical study. A single DINOv3-pretrained ConvNeXt trained with a tuned recipe achieves 88.19 mAP on VeRi-Wild Small and 77.47 mAP on VeRi-Wild Large using visual cues alone, matching the strongest protocol-verified metadata-dependent multi-branch baseline. Applying training-free re-ranking further improves performance to 92.38 and 83.68 mAP, respectively. Using this strong baseline together with retrieval-level branch diagnostics, we evaluate whether increasing representational diversity still provides measurable gains. Across both benchmarks, concatenating multiple branches built on a shared backbone changes the best single-branch performance by less than one mAP point while increasing the embedding dimension by 4x, and the resulting representation has an effective rank close to the original feature dimension. We further study cross-backbone fusion using an asymmetric frozen-anchor strategy to combine ConvNeXt and Vision Transformer representations. Despite these favorable conditions, Transformer branches consistently remain 13-15 mAP below the ConvNeXt backbone, and paired per-query bootstrap analysis estimates the largest observed fusion gain to be only +0.11 mAP (95% confidence interval). Our results suggest that, under the evaluated setting, improving a single strong foundation-model backbone together with retrieval-stage re-ranking is more effective than increasing architectural complexity through additional branches or heterogeneous backbones. We restrict our conclusions to single-seed training and one family of foundation models and discuss conditions under which these observations may not hold.

Figures

Figures reproduced from arXiv: 2607.22068 by the authors.

Figure 1
Figure 1. FIGURE 1 [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 7 linked inside Pith

  1. [1]

    VERI-Wild: A large dataset and a new method for vehicle re-identification in the wild,

    Y . Lou, Y . Bai, J. Liu, S. Wang, and L.-Y . Duan, “VERI-Wild: A large dataset and a new method for vehicle re-identification in the wild,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2019, pp. 3235–3243

  2. [2]

    A deep learning-based approach to progressive vehicle re-identification for urban surveillance,

    X. Liu, W. Liu, T. Mei, and H. Ma, “A deep learning-based approach to progressive vehicle re-identification for urban surveillance,” inProc. European Conf. Computer Vision (ECCV), 2016, pp. 869–884

  3. [3]

    A comprehensive survey on deep-learning-based vehicle re-identification: Models, data sets and challenges,

    A. Amiri, A. Kaya, and A. S. Kec ¸eli, “A comprehensive survey on deep-learning-based vehicle re-identification: Models, data sets and challenges,”arXiv:2401.10643, 2024

  4. [4]

    Learning discrimi- native features with multiple granularities for person re-identification,

    G. Wang, Y . Yuan, X. Chen, J. Li, and X. Zhou, “Learning discrimi- native features with multiple granularities for person re-identification,” inProc. ACM Int. Conf. Multimedia (MM), 2018, pp. 274–282

  5. [5]

    Strength in diversity: Multi- branch representation learning for vehicle re-identification,

    E. Almeida, B. Silva, and J. Batista, “Strength in diversity: Multi- branch representation learning for vehicle re-identification,” inProc. IEEE Int. Conf. Intelligent Transportation Systems (ITSC), 2023

  6. [6]

    TransReID: Transformer-based object re-identification,

    S. He, H. Luo, P. Wang, F. Wang, H. Li, and W. Jiang, “TransReID: Transformer-based object re-identification,” inProc. IEEE/CVF Int. Conf. Computer Vision (ICCV), 2021, pp. 15 013–15 022

  7. [7]

    Unity is strength: Unifying convolutional and transformeral features for better person re- identification,

    Y . Wang, P. Zhang, X. Liu, Z. Tu, and H. Lu, “Unity is strength: Unifying convolutional and transformeral features for better person re- identification,”IEEE Trans. Intelligent Transportation Systems, 2024

  8. [8]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquabet al., “DINOv3,”arXiv:2508.10104, Meta AI Research, 2025

Show all 31 references
  1. [9]

    CLIP-ReID: Exploiting vision-language model for image re-identification without concrete text labels,

    S. Li, L. Sun, and Q. Li, “CLIP-ReID: Exploiting vision-language model for image re-identification without concrete text labels,” inProc. AAAI Conf. Artificial Intelligence, 2023

  2. [10]

    Fine- tuning can distort pretrained features and underperform out-of- distribution,

    A. Kumar, A. Raghunathan, R. Jones, T. Ma, and P. Liang, “Fine- tuning can distort pretrained features and underperform out-of- distribution,” inProc. Int. Conf. Learning Representations (ICLR), 2022

  3. [11]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11 976–11 986

  4. [12]

    ResNet strikes back: An improved training procedure in timm,

    R. Wightman, H. Touvron, and H. J ´egou, “ResNet strikes back: An improved training procedure in timm,”arXiv:2110.00476, 2021

  5. [13]

    Bag of tricks and a strong baseline for deep person re-identification,

    H. Luo, Y . Gu, X. Liao, S. Lai, and W. Jiang, “Bag of tricks and a strong baseline for deep person re-identification,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

  6. [14]

    Deep ensembles: A loss landscape perspective,

    S. Fort, H. Hu, and B. Lakshminarayanan, “Deep ensembles: A loss landscape perspective,”arXiv:1912.02757, 2019

  7. [15]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” in Advances in Neural Information Processing Systems (NeurIPS), 2017

  8. [16]

    Do vision transformers see like convolutional neural networks?

    M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy, “Do vision transformers see like convolutional neural networks?” in Advances in Neural Information Processing Systems (NeurIPS), 2021

  9. [17]

    Similarity of neural network representations revisited,

    S. Kornblith, M. Norouzi, H. Lee, and G. Hinton, “Similarity of neural network representations revisited,” inProc. Int. Conf. Machine Learning (ICML), 2019

  10. [18]

    Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth,

    T. Nguyen, M. Raghu, and S. Kornblith, “Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth,” inProc. Int. Conf. Learning Representa- tions (ICLR), 2021

  11. [19]

    Re-ranking person re- identification with k-reciprocal encoding,

    Z. Zhong, L. Zheng, D. Cao, and S. Li, “Re-ranking person re- identification with k-reciprocal encoding,” inProc. IEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2017, pp. 1318–1327

  12. [20]

    Deep relative distance learning: Tell the difference between similar vehicles,

    H. Liu, Y . Tian, Y . Wang, L. Pang, and T. Huang, “Deep relative distance learning: Tell the difference between similar vehicles,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2167–2175

  13. [21]

    CLIP-SENet: CLIP- based semantic enhancement network for vehicle re-identification,

    L. Lu, Z. Fu, D. Chu, W. Wang, and B. Xu, “CLIP-SENet: CLIP- based semantic enhancement network for vehicle re-identification,” arXiv:2502.16815, 2025

  14. [22]

    BEiT: BERT pre-training of image transformers,

    H. Bao, L. Dong, S. Piao, and F. Wei, “BEiT: BERT pre-training of image transformers,” inProc. Int. Conf. Learning Representations (ICLR), 2022

  15. [23]

    Circle loss: A unified perspective of pair similarity optimization,

    Y . Sun, C. Cheng, Y . Zhang, C. Zhang, L. Zheng, Z. Wang, and Y . Wei, “Circle loss: A unified perspective of pair similarity optimization,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2020, pp. 6398–6407

  16. [24]

    A discriminative feature learning approach for deep face recognition,

    Y . Wen, K. Zhang, Z. Li, and Y . Qiao, “A discriminative feature learning approach for deep face recognition,” inProc. European Conf. Computer Vision (ECCV), 2016, pp. 499–515

  17. [25]

    In defense of the triplet loss for person re-identification,

    A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,”arXiv:1703.07737, 2017

  18. [26]

    Visualizing and understanding convolu- tional networks,

    M. D. Zeiler and R. Fergus, “Visualizing and understanding convolu- tional networks,” inProc. European Conf. Computer Vision (ECCV), 2014, pp. 818–833

  19. [27]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inProc. Int. Conf. Learning Representations (ICLR), 2021

  20. [28]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inProc. Int. Conf. Learning Representations (ICLR), 2022

  21. [29]

    Param- eter efficient fine-tuning of self-supervised ViTs without catastrophic forgetting,

    R. Akbarian Bafghi, N. Harilal, C. Monteleoni, and M. Raissi, “Param- eter efficient fine-tuning of self-supervised ViTs without catastrophic forgetting,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition Workshops (CVPRW), 2024, arXiv:2404.17245

  22. [30]

    Fine-tuning CNN image retrieval with no human annotation,

    F. Radenovi ´c, G. Tolias, and O. Chum, “Fine-tuning CNN image retrieval with no human annotation,”IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 41, no. 7, pp. 1655–1668, 2019

  23. [31]

    Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline),

    Y . Sun, L. Zheng, Y . Yang, Q. Tian, and S. Wang, “Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline),” inProc. European Conf. Computer Vision (ECCV), 2018, pp. 480–496. VOLUME , 9

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.