Pith. sign in

REVIEW 4 major objections 5 minor 38 references

DART$^3$: Leveraging Distance for Test Time Adaptation in Person Re-Identification

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Person re-identification models can be adapted to unseen cameras by learning per-camera scale and shift parameters with a top-k distance loss, which the paper shows outperforms entropy-based test-time adaptation on three benchmarks.

desk verdict A promising distance-based TTA for ReID whose practical black-box variant and honest ablation are undermined by a theory the authors admit is unrealistic and by thin, variance-free empirical margins. read the letter →

arxiv 2505.18337 v1 pith:USVCOFCU submitted 2025-05-23 cs.CV

classification cs.CV
keywords personre-identificationtest-timeadaptationcamerabiasdistance-basedretrievalobjectivescale-shiftparameterblack-boxdomainshift
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

Person re-identification models embed images into vectors that cluster by the camera that captured them, not by the person's identity, and this camera bias worsens when new cameras are added to a surveillance network. The paper claims that the bias can be modeled as a per-camera affine warp $z_i = \alpha_{c_i} z^*_i + \beta_{c_i}$ of an ideal identity representation $z^*_i$, and that a test-time adaptation objective based on top-$k$ Euclidean distances to the nearest gallery features can learn the scale and shift parameters that undo the warp. The proposed method, DART$^3$, requires no source data, no architectural changes, and no retraining, and its LITE variant treats the source model as a black box, adapting only about 30k–50k external parameters. On MSMT17, Duke-MTMC, and Market-1501, with CLIP-ReID and TransReID backbones, the paper reports consistent gains over entropy-based TTA baselines and camera normalization, particularly when the test set contains cameras unseen during training. If these results hold, distance-based tuning gives surveillance systems a deployable, source-free way to keep person retrieval accurate as camera networks grow.

What carries the argument

The machinery is the per-camera affine bias model $z_i = \alpha_{c_i} z^*_i + \beta_{c_i}$ paired with learnable external dictionaries $M(c_i)$ and $\Sigma(c_i)$, initialized as per-camera means and standard deviations and updated only for the query features while the gallery normalization stays frozen. The adaptation loss is the mean of a top-$k$ masked softmax over negative Euclidean distances between scaled-shifted queries and normalized gallery features, $\mathcal{L} = \frac{1}{B}\sum_i\sum_j H(\Delta)(i,j)\,M_{\mathrm{top-}k}(i,j)$, which the paper shows correlates with error rate more stably than entropy or cosine distance. This turns camera bias into a small, invertible per-camera transformation that can be tuned without source data, architectural changes, or, in the LITE variant, any access to the model weights.

What would settle it

Split a test set from a fixed camera network into two temporal halves, recompute the per-camera mean and standard deviation on each half, and run DART$^3$ with each initialization: if the gains over static camera normalization shrink or reverse whenever the two halves' statistics differ, the camera-constant affine-bias assumption is falsified. A complementary experiment would inject a synthetic non-affine or identity-dependent bias into features and check whether DART$^3$ can still recover the ideal representation; the paper's own derivation predicts it cannot.

Watch

Extended reading notes

Core claim

The central discovery is that Euclidean distance, rather than entropy, is the right test-time signal for person re-identification. The paper shows that for a CLIP-ReID model under unseen-camera shift, the error rate rises more monotonically with the Euclidean distance to the nearest gallery match than with cosine distance or prediction entropy, so a softmax over negative top-$k$ Euclidean distances provides a stable learning objective. Combined with external per-camera scale-shift parameters $M(c_i)$, $\Sigma(c_i)$ initialized as per-camera feature mean and standard deviation and updated only on the query side, this objective is claimed to remove a large part of the camera-induced bias. Empirically, DART$^3$ and DART$^3$ LITE outperform TENT, TEMP, and camera normalization across three benchmarks and two backbones, with the largest gains in the new-camera-in-existing-network setting.

Load-bearing premise

The method assumes each camera's bias is a single, constant per-camera scale-and-shift applied to an ideal feature representation that is already mean-zero with unit variance, so the per-camera mean and standard deviation really are the bias parameters; if bias varies by identity or time or is nonlinear, the derivation collapses.

Editorial extensions

If this is right

  • Because DART$^3$ LITE adapts only external scale-shift parameters (~30k–50k), the source model can remain a true black box, so adaptation can be applied to third-party or privacy-protected ReID models without weight access.
  • The reported results imply that entropy minimization, successful in classification TTA, is a poor objective for retrieval tasks whose predictions depend on distances between unbounded feature vectors.
  • The ablation shows the distance loss alone hurts performance and the per-camera scale-shift parameters are the component that removes camera bias, confirming the affine-bias model's role.
  • Performance improves with additional optimization steps in the non-episodic setting, so the method is suited to online adaptation as test batches arrive from new cameras.
  • With grounding samples from seen cameras, performance rises further, indicating that some availability of source-domain data can be exploited when present, without breaking the core source-free design.

Reading between the lines

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

  • If the affine per-camera model generalizes, the same scale-shift plus top-$k$ distance prescription could transfer to other retrieval settings with known capture-site metadata, such as vehicle re-identification or camera-trap animal matching.
  • The paper's own noise simulation (Supp. B.4) suggests a clearest test of the method's boundary: deliberately corrupt the per-camera statistics at initialization and measure how quickly gains over static normalization vanish.
  • Because Table 3 shows DART$^3$ does not improve source-domain performance, a practical deployment would want a domain-shift detector that enables adaptation only when unseen cameras appear.
  • The initialization's dependence on per-camera statistics implies that long-running cameras with seasonal or lighting drift may violate the constancy assumption; testing on temporally extended feeds would probe that boundary.
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 DART3, a test-time adaptation (TTA) method for person re-identification (ReID) under camera-induced domain shift. The method combines a distance-based objective that minimizes the top-k Euclidean distances between query features and normalized gallery features (Eqs. 1–2) with learnable per-camera scale-shift parameters (Eqs. 3–5). The scale-shift parameters are initialized using per-camera feature means and standard deviations, following the camera-normalization idea of [24], and only the query-side parameters are trained while the gallery-side parameters are frozen. A 'LITE' variant adapts only the external scale-shift parameters, making the source model fully black-box. The method is evaluated on MSMT17, Duke-MTMC, and Market-1501 under two protocols: adding unseen cameras to an existing network (Table 1) and deploying on a completely new camera network (Table 2). An ablation study (Table 4) shows that the distance loss alone slightly hurts performance, the camera-statistics initialization alone matches [24], and only their combination gives the reported improvement. The authors conclude that DART3 consistently outperforms entropy-based TTA baselines while requiring no source data or architectural changes.

Significance. If the empirical claims hold, DART3 would be a useful contribution: it introduces a retrieval-oriented TTA objective that is better matched to ReID than classification entropy, it can operate in a black-box mode with only a small number of external parameters, and the ablation honestly isolates the contribution of each component. The paper also makes a falsifiable claim about the relationship between retrieval error and Euclidean nearest-neighbor distance. However, the current evidence is not yet sufficient to support the central claim of consistent improvement: the reported gains are small in several settings, the paper provides no variance or significance estimates, and hyperparameters are selected on the same target-domain protocol used for the main tables. The theoretical motivation is acknowledged by the authors themselves to rest on unrealistic assumptions. These issues are addressable with additional experiments and a tightened experimental protocol, so the work is potentially publishable after major revision.

major comments (4)
  1. [Table 1 / Sec. 4.2] The central claim that DART3 consistently outperforms the baselines is not yet supported by the reported statistics. All entries in Tables 1 and 2 appear to come from a single run, with no standard deviation, multiple seeds, or significance testing. In several cells the margin over Camera Normalization [24] is very small or negative: for TransReID on Market-1501, DART3 has mAP 85.8 versus 86.6 for Norm, and for CLIP-ReID on Market-1501, DART3 has mAP 86.7 versus 86.6 for Norm. Given that the headline gains are only 1–4 mAP in most settings, the authors should report results over multiple seeds with means and standard deviations, and where possible paired comparisons, before claiming consistent superiority.
  2. [Sec. 4.4 / Fig. 5a] The hyperparameters are selected on the same target-domain protocol used for the main evaluation. The ablation study uses 'an unseen camera as the target data' from MSMT17 with the same CLIP-ReID backbone and the same 5-camera training split as Table 1, and from this the authors select k = 3, tau = 100–200, lr = 1e-4, and one optimization step. Since Table 1 is then evaluated on the remaining unseen cameras of the same dataset, there is a risk of test-set hyperparameter fitting. The authors should either fix the hyperparameters based only on source-domain development data, use a separate validation camera or dataset that is excluded from the reported tables, or provide evidence that the results are insensitive to these choices across the evaluation settings.
  3. [Supp. B.4 / Sec. 3.2] The theoretical justification for the camera-normalization initialization is explicitly acknowledged by the authors to rest on unrealistic assumptions. Equations (15)–(17) derive the equivalence between the normalized biased features and the normalized unbiased features only under the affine model z_i = alpha_c z*_i + beta_c and under the standardization assumption z*_i ~ N(0, 1). The paper states in Supp. B.4 that 'for real world data, these assumptions are unrealistic.' Consequently, Eq. (17) cannot be presented as a proof that per-camera mean/std normalization removes camera bias; it is at best a heuristic initialization. The statement in Sec. 4.4 that the ablation 'validates our mathematical model' is too strong. The authors should either temper the theoretical claims to describe a motivating approximation, or provide a direct empirical test of the affine-bias model, such as examining whether per-camera residual biases after normalization are affine in the unbiased features.
  4. [Table 3 / Sec. 5] The source-domain results weaken the claim that DART3 is generally beneficial. In Table 3, DART3 is sometimes worse than No-adapt, for example TransReID on MSMT17 (mAP 61.9 versus 66.4), and the paper's own Limitations section acknowledges that DART3 is unable to improve source-distribution performance. This is not by itself a fatal issue, since the method targets out-of-domain camera shifts, but the abstract's phrase 'consistently outperforms' should be restricted to the target-domain settings, and the source-domain behavior should be discussed as a failure mode rather than merely listed as a limitation.
minor comments (5)
  1. [Eq. (1)] The definition of the top-k mask is confusing: the condition refers to 'Δ_soft(i, j)' being among the top-k least values 'within H(Δ)(i, ...)', which mixes two different tensors. Please clarify whether the mask is computed from Δ_soft or from H(Δ), and specify how ties are handled.
  2. [Algorithm 2 / Supp. B.5] The pseudocode says 'Update learnable parameters i.e. batch-norm parameters μ, σ∈ θ, M_q and Σ_q', which contains a typographical error (σ∈ θ should presumably be σ, θ or similar). Please also state explicitly which parameters are updated in the LITE variant.
  3. [Table 2] Table 2 appears misformatted, with duplicated entries in several rows (e.g., the TransReID Upper Bound row repeats the same three numbers interleaved). The table should be reformatted so each column corresponds to one source-target pair.
  4. [Sec. 4.4 / Fig. 5d] The grounding-sample experiment is presented only as a figure, with no numerical table or protocol description. Please report the exact setup, the number of grounding samples, and quantitative results for all methods, or move this experiment to the supplementary material with full details.
  5. [Implementation Details] The paper does not state whether the reported results use the non-episodic mode for all tables and whether the one-step-per-batch default is used throughout. Please clarify this for each experiment, since the number of optimization steps has a large effect on behavior (Fig. 5b–5c).

Circularity Check

1 steps flagged · score 2.0 of 10

Minor definitional circularity in the formal motivation for the distance objective; the central empirical claim is not circular.

  1. self definitional [Supplementary Material B.2, Proposition 1 / Corollary 2]
    "Proposition 1: Let xq and xg be two feature vectors in a Euclidean space Rd, representing a query and a gallery sample respectively, both belonging to the same class. Let u : R≥0 → R≥0 be a function that models the uncertainty in classification as a function of the Euclidean distance d(xq, xg) = ∥xq − xg∥2. Then, it is intuitive to assume that u satisfies the following properties: ... 2. Monotonicity: u is monotonically increasing, i.e., if d1 < d2, then u(d1) < u(d2). ... Corollary 2 (Uncertainty grows with distance)."

    Corollary 2 is a verbatim restatement of the monotonicity assumption already listed as Property 2 of the same Proposition. No derivation is supplied; the conclusion 'uncertainty grows with distance' is true by definition of the assumed function u, not as an independent mathematical result. This formal framing is used to motivate the distance-based loss, although the paper also cites empirical evidence (Figure 7, Table 7), so the circular element is limited to the presentation and does not by itself carry the paper's central claim.

full rationale

The paper's central claim—that DART3 and DART3 LITE improve test-time person ReID under camera shift—is supported by external benchmark evaluations (Tables 1 and 2) and by an ablation (Table 4) showing that the distance loss alone does not improve performance without the normalized scale-shift initialization. The per-camera normalization equivalence in Eq. 17 is a conditional algebraic consequence of the explicitly stated affine bias model in Eq. 12, and the authors themselves disclaim its real-world validity: 'the reliability of the Normalization algorithm rests entirely on the two assumptions: (1) z*_i ~ N(0,1) and (2) alpha and beta are constants with respect to camera. We argue that for real world data, these assumptions are unrealistic.' Since the assumptions are stated and the equivalence is used only for initialization, this is not a hidden circularity. The choice of Euclidean distance over entropy is supported by the empirical correlation in Figure 2/Figure 7 rather than by the Proposition alone. The main non-circularity concern is that hyperparameters (k, tau, lr) are selected using an unseen camera from the same benchmark that later appears in Table 1; this is a test-set-selection / external-validity risk, not a derivation-level circularity and is partially acknowledged in the Limitations section. No load-bearing self-citation chain or uniqueness argument appears, and the reported gains are not forced by construction. Score 2 reflects only the minor definitional framing in Supplementary B.2.

Assumptions & free parameters 5 free parameters · 6 assumptions · 3 invented entities

The central claim rests on an assumed affine bias model, on a standardization assumption the authors label unrealistic, and on a monotonic distance-uncertainty axiom supported only by a qualitative curve. The method adds per-camera learnable vectors initialized from target-data statistics plus several tuned hyperparameters (tau, k, lr, steps). All introduced entities (per-camera parameters, the latent z*, grounding samples) are validated only through the paper's own experiments, not through any external falsifiable handle.

free parameters (5)
  • Softmax temperature tau = 100 for CLIP-ReID, 200 for TransReID
    Chosen per backbone; the paper states that lower values cause gradients to vanish due to precision limits (Sec. 4.1), so the operating point is a manual choice rather than a derived one.
  • Top-k value k = 3 (default), best in range 1 to 5
    Selected from ablation sweeps in Figure 5a and Supp. C.2; the paper notes k should be small when temperature is low, making the pair a tuned operating point.
  • Learning rate = 1e-4
    Reported as a fixed choice with no tuning rule or schedule for different targets (Sec. 4.1).
  • Optimization steps per batch = 1 (default)
    The paper reports that performance keeps improving with more steps (Fig. 5b/5c) yet reports all tables with one step, a manually chosen operating point.
  • Per-camera mean and standard deviation (mu_c, sigma_c) = estimated from target features for each camera
    These statistics initialize the external parameters and implicitly estimate the assumed bias terms alpha and beta; the paper's own simulation (Fig. 8) shows such estimates are sensitive to noise.
assumptions (6)
  • domain assumption Camera bias is affine and camera-constant: z_i = α_{c_i} z*_i + β_{c_i} (Eq. 3 in main text, Eq. 12 in Supp.)
    The scale-shift machinery and the normalization identity (Eqs. 15-17) follow from this model; the paper presents it as an assumed general case of additive and multiplicative bias (Supp. A).
  • domain assumption Unbiased features are standardized, effectively N(0,1), so per-camera mean and std equal the bias parameters
    Supp. B.4 states the reliability of the normalization rests on this assumption and that it is unrealistic for real data; the direct recovery of alpha and beta in Eqs. 15-16 depends on it.
  • domain assumption Retrieval uncertainty grows monotonically with Euclidean distance (Proposition 1, Supp. B.2)
    Stated as intuitive and supported only by the qualitative curve in Figure 2, which lacks binning details and error bars; the distance-based loss is built on this premise.
  • ad hoc to paper The top-k nearest gallery points to a query carry a useful signal for adaptation
    The objective pulls queries toward currently nearest gallery features; the paper's ablation shows this alone hurts performance (41.7 vs 42.1 mAP for no-adapt), so the assumption only holds inside the full combination.
  • ad hoc to paper Freezing the gallery-side parameters (M_g, Sigma_g) prevents trivial collapse
    Stated as a design choice in Sec. 3.2 without analysis of the collapse mode it avoids.
  • domain assumption Camera IDs are known for all query and gallery images at test time
    Per-camera dictionaries are keyed by camera ID (Sec. 3), which may be unavailable or unreliable in real deployments with added cameras.
invented entities (3)
  • External learnable per-camera scale-shift parameters (M_q, Sigma_q, M_g, Sigma_g)
    purpose: Estimate and undo the assumed per-camera affine bias on query and gallery features at test time
    The only evidence is the paper's own benchmark experiments; no prediction outside the paper validates these entities separately.
  • True unbiased representation z*_i
    purpose: The bias-free feature each image would have under the affine model (Sec. 3.2)
    Never observed directly; it exists only inside the assumed bias model and is estimated via normalization rather than validated against any external quantity.
  • Grounding samples (source-domain images inserted into test batches)
    purpose: Boost adaptation performance when source data happens to be available at test time (Fig. 5d)
    The paper concedes this breaks the source-free setup and justifies it only by the observed performance increase.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DART$^3$: Leveraging Distance for Test Time Adaptation in Person Re-Identification." pith.science (2026). https://pith.science/paper/USVCOFCU

@misc{pith2026250518337,
  author       = {Pith},
  title        = {Pith review of: DART$^3$: Leveraging Distance for Test Time Adaptation in Person Re-Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/USVCOFCU}},
  note         = {Machine review of arXiv:2505.18337}
}
abstract

Person re-identification (ReID) models are known to suffer from camera bias, where learned representations cluster according to camera viewpoints rather than identity, leading to significant performance degradation under (inter-camera) domain shifts in real-world surveillance systems when new cameras are added to camera networks. State-of-the-art test-time adaptation (TTA) methods, largely designed for classification tasks, rely on classification entropy-based objectives that fail to generalize well to ReID, thus making them unsuitable for tackling camera bias. In this paper, we introduce DART$^3$, a TTA framework specifically designed to mitigate camera-induced domain shifts in person ReID. DART$^3$ (Distance-Aware Retrieval Tuning at Test Time) leverages a distance-based objective that aligns better with image retrieval tasks like ReID by exploiting the correlation between nearest-neighbor distance and prediction error. Unlike prior ReID-specific domain adaptation methods, DART$^3$ requires no source data, architectural modifications, or retraining, and can be deployed in both fully black-box and hybrid settings. Empirical evaluations on multiple ReID benchmarks indicate that DART$^3$ and DART$^3$ LITE, a lightweight alternative to the approach, consistently outperforms state-of-the-art TTA baselines, making for a viable option to online learning to mitigate the adverse effects of camera bias.

Figures

Figures reproduced from arXiv: 2505.18337 by the authors.

Figure 1
Figure 1. A failure case where the source model misidentifies a query image due to camera [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Change in error rate vs nearest Eu￾clidean and Cosine distance, and entropy, over a range of values. The error rate for Euclidean distance grows more uniformly over the range. In contrast, cosine distance grows abruptly for a high value, and with entropy the error rate be￾haves unpredictably [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. We present DART3 , a test time adaptation pipeline designed to mitigate camera bias in ReID models when exposed to unseen camera domains. We hypothesize that a true unbiased representation z ∗ i exists and can be estimated by scale and shift parameters. For the adaptation, these parameters can be initialized as the mean and standard deviation of features accumulated for a specific camera. Finally we show two variant… view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: Trends in performance (mAP) with respect to the value of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: While the surrounding clusters are formed based on the person ID, as in the image, we [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 6
Figure 6. Figure 6: UMAP representation of two person ID clusters from Market-1501. Features are from [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Change in error rate vs nearest Eu￾clidean and Cosine distance, and entropy, over a range of values. The error rate for Euclidean distance grows more uniformly over the range. In contrast, cosine distance grows abruptly for a high value, and with entropy the error rate…
Figure 8
Figure 8. Figure 8: We compare (1) Estimate hit rate (i.e. frequency of correct estimates within 0.01 error [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Performance comparison (mAP) for each camera in MSMT17 dataset. Greyed out camera [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Trends in performance (mAP) with respect to the value of (a) [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Comparison of performance with respect to number of samples used per camera to estimate [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages

  1. [24]

    Exploring the Camera Bias of Person Re-identification

    Myungseo Song, Jin-Woo Park, and Jong-Seok Lee. Exploring the camera bias of person re-identification. arXiv preprint arXiv:2502.10195, 2025

  2. [1]

    Test-time similarity modification for person re-identification toward temporal distribution shift

    Kazuki Adachi, Shohei Enomoto, Taku Sasaki, and Shin’Ya Yamaguchi. Test-time similarity modification for person re-identification toward temporal distribution shift. In2024 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2024

  3. [2]

    Anchor-reid: A test time adaptation for person re-identification

    Mohammed Almansoori, Mustansar Fiaz, and Hisham Cholakkal. Anchor-reid: A test time adaptation for person re-identification. In Scandinavian Conference on Image Analysis, pages 599–612. Springer, 2023

  4. [3]

    Parameter-free online test-time adaptation

    Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344–8353, 2022

  5. [4]

    Contrastive test-time adaptation

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 295–305, 2022

  6. [5]

    Ice: Inter-instance contrastive encoding for unsupervised person re-identification

    Hao Chen, Benoit Lagadec, and Francois Bremond. Ice: Inter-instance contrastive encoding for unsupervised person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14960–14969, 2021

  7. [6]

    Part-based pseudo label re- finement for unsupervised person re-identification

    Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label re- finement for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7308–7318, 2022

  8. [7]

    Test time adaptation via conjugate pseudo-labels

    Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems, 35:6204–6218, 2022

Show all 38 references
  1. [8]

    1st place solution to visda-2020: Bias elimination for domain adaptive pedestrian re-identification

    Jianyang Gu, Hao Luo, Weihua Chen, Yiqi Jiang, Yuqi Zhang, Shuting He, Fan Wang, Hao Li, and Wei Jiang. 1st place solution to visda-2020: Bias elimination for domain adaptive pedestrian re-identification. arXiv preprint arXiv:2012.13498, 2020

  2. [9]

    Clipartt: Light-weight adaptation of clip to new domains at test time

    Gustavo Adolfo Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Clipartt: Light-weight adaptation of clip to new domains at test time. arXiv preprint arXiv:2405.00754, 2024

  3. [10]

    Generalizable person re- identification via self-supervised batch norm test-time adaption

    Ke Han, Chenyang Si, Yan Huang, Liang Wang, and Tieniu Tan. Generalizable person re- identification via self-supervised batch norm test-time adaption. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 817–825, 2022. 10

  4. [11]

    Transreid: Transformer- based object re-identification

    Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer- based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013–15022, 2021

  5. [12]

    Self domain adapted network

    Yufan He, Aaron Carass, Lianrui Zuo, Blake E Dewey, and Jerry L Prince. Self domain adapted network. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 437–446. ...

  6. [13]

    Test-time classifier adjustment module for model-agnostic domain generalization

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427–2440, 2021

  7. [14]

    Cloth-changing person re-identification from a single image with gait prediction and regularization

    Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In Proceedings of the IEEE/CVF conference on c...

  8. [15]

    Test-time adaptable neural networks for robust medical image segmentation

    Neerav Karani, Ertunc Erdil, Krishna Chaitanya, and Ender Konukoglu. Test-time adaptable neural networks for robust medical image segmentation. Medical Image Analysis, 68:101907, 2021

  9. [16]

    Camera-driven representation learning for unsupervised domain adaptive person re- identification

    Geon Lee, Sanghoon Lee, Dohyung Kim, Younghoon Shin, Yongsang Yoon, and Bumsub Ham. Camera-driven representation learning for unsupervised domain adaptive person re- identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11453–11462, 2023

  10. [17]

    Clip-reid: exploiting vision-language model for image re- identification without concrete text labels

    Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re- identification without concrete text labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1405–1413, 2023

  11. [18]

    Generalizing person re-identification by camera-aware invariance learning and cross-domain mixup

    Chuanchen Luo, Chunfeng Song, and Zhaoxiang Zhang. Generalizing person re-identification by camera-aware invariance learning and cross-domain mixup. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 224–241....

  12. [19]

    Test-time adaptation with calibration of medical image classification nets for label distribution shift

    Wenao Ma, Cheng Chen, Shuang Zheng, Jing Qin, Huimao Zhang, and Qi Dou. Test-time adaptation with calibration of medical image classification nets for label distribution shift. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 313...

  13. [20]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International confer- ence on machine learning, pages 16888–16905. PMLR, 2022

  14. [21]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

  15. [22]

    Performance measures and a data set for multi-target, multi-camera tracking

    Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17–35. Springer, 2016

  16. [23]

    Body part-based rep- resentation learning for occluded person re-identification

    Vladimir Somers, Christophe De Vleeschouwer, and Alexandre Alahi. Body part-based rep- resentation learning for occluded person re-identification. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1613–1623, 2023

  17. [25]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 11

  18. [26]

    Towards understanding gd with hard and conjugate pseudo-labels for test-time adaptation

    Jun-Kun Wang and Andre Wibisono. Towards understanding gd with hard and conjugate pseudo-labels for test-time adaptation. arXiv preprint arXiv:2210.10019, 2022

  19. [27]

    Camera- aware proxies for unsupervised person re-identification

    Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera- aware proxies for unsupervised person re-identification. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 2764–2772, 2021

  20. [28]

    Person transfer gan to bridge domain gap for person re-identification

    Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018

  21. [29]

    Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning

    Zesen Wu and Mang Ye. Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9548–9558, 2023

  22. [30]

    Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images

    Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41(12):3575–3586, 2022

  23. [31]

    Visual domain bridge: A source-free domain adaptation for cross-domain few-shot learning

    Moslem Yazdanpanah and Parham Moradi. Visual domain bridge: A source-free domain adaptation for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2868–2877, 2022

  24. [32]

    Test-time batch statistics calibration for covariate shift

    Fuming You, Jingjing Li, and Zhou Zhao. Test-time batch statistics calibration for covariate shift. arXiv preprint arXiv:2110.04065, 2021

  25. [33]

    Unsupervised domain adaptation for person re-identification via heterogeneous graph alignment

    Minying Zhang, Kai Liu, Yidong Li, Shihui Guo, Hongtao Duan, Yimin Long, and Yi Jin. Unsupervised domain adaptation for person re-identification via heterogeneous graph alignment. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3360–3368, 2021

  26. [34]

    Scalable person re-identification: A benchmark

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015

  27. [35]

    Bayesian adaptation for covariate shift

    Aurick Zhou and Sergey Levine. Bayesian adaptation for covariate shift. Advances in neural information processing systems, 34:914–927, 2021

  28. [36]

    Rethinking the distribution gap of person re-identification with camera-based batch normalization

    Zijie Zhuang, Longhui Wei, Lingxi Xie, Tianyu Zhang, Hengheng Zhang, Haozhe Wu, Haizhou Ai, and Qi Tian. Rethinking the distribution gap of person re-identification with camera-based batch normalization. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, Augu...

  29. [37]

    Boundary condition: u(0) = 0 Interpretation: If the query and gallery embeddings are identical, the prediction is made with absolute certainty

  30. [38]

    No-adapt

    Monotonicity: u is monotonically increasing, i.e., if d1 < d2, then u(d1) < u(d2). Interpretation: As the distance between query and gallery increases, the uncertainty in prediction also increases. 15 Corollary 1 (Certainty at zero distance). If xq = xg, then d(xq, xg) = 0 and...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.