PaperlayerView exact revision on arXiv

Source-linked AI summary

MINER: Modeling Inter- and Intra-Person Variance for Reidentification

Oron Nir, Eliyahu Strugo, Ariel Shamir

arXiv:2610.28941v2cs.CV

TL;DR

Clothing changes can dominate the appearance signal in CC-ReID, while prior methods often leave within-person cross-outfit variation as a byproduct of metric learning. MINER combines stochastic hard-example mining with adaptive allocation of training batches to balance intra- and inter-person variance. It improves retrieval across CC-ReID benchmarks, achieves state-of-the-art performance on four datasets, and ranks top-3 in all TVRID 2026 tracks.

  • Problem

    Prior CC-ReID methods emphasize inter-person separation but treat within-person cross-outfit variation as a byproduct rather than a first-class objective.

    Source: §1 · p. 1

  • Method

    MINER combines stochastic hard mining of similar-identity negatives and dissimilar cross-clothing positives with adaptive batch composition that balances intra- and inter-person variance.

    Source: §1 · p. 1, §3.3 · p. 4

  • Results

    MINER improves retrieval across CC-ReID benchmarks, achieves state-of-the-art performance on four datasets, and ranks top-3 in all TVRID 2026 tracks.

    Source: §1 · p. 1, §7 · p. 8

  • Takeaways & Limitations

    The results support variance-aware training as a principled path toward robust identity representations under appearance shift.

    Source: §7 · p. 8

    MINER assumes vanilla-pretrained weights, requires clothing-ID labels and at least two outfits per identity, and limited clothing diversity weakens hard-positive mining.

    Source: §6 · p. 8

Abstract

from arXiv · show

Clothes-Changing Person Re-Identification (CC-ReID) remains challenging due to drastic appearance shifts across outfits. Most prior works devise biometric signal-extraction heuristics to enhance inter-person variance, e.g., using segmentation masks, while leaving intra-person variance uncontrolled. We present MINER, a simple model-agnostic framework that suppresses garment bias by contrasting hard examples and adaptively balancing dataset-dependent intra- and inter-person variance. Extensive experiments show that MINER improves retrieval across CC-ReID benchmarks, achieving state-of-the-art performance on four datasets and ranking among the leaders in all three tracks of the open TVRID RGB-Depth ReID challenge.

1. Introduction

CC-ReID must identify people despite clothing changes that can dominate visual appearance. MINER addresses a gap in prior work by explicitly balancing within-person cross-outfit invariance and between-person discrimination through stochastic hard mining and adaptive batch composition.

1. Introduction

  • CC-ReID must handle clothing-induced appearance changes that often dominate the visual signal, requiring identity cues that remain invariant to garments.The task retrieves the same person across cameras despite outfit changes over time.Source: §1 · p. 1
  • Prior methods emphasize inter-person separation, while treating within-person differences across outfits as a byproduct of metric learning rather than a direct objective.These approaches seek to suppress clothing features and enhance biometric cues.Source: §1 · p. 1
  • MINER combines stochastic hard mining with multi-positive contrastive learning and an online controller that adapts the balance between inter-person discrimination and cross-clothing invariance.It soft-samples visually similar identities as hard negatives and maximally dissimilar cross-camera outfits as hard positives; the controller adapts the trade-off from gradient variance without grid search or extra forward passes.Source: §1 · p. 1
  • MINER improves state-of-the-art backbones on four CC-ReID benchmarks and ranks top-3 in all TVRID 2026 tracks.The benchmarks are PRCC, LTCC, Celeb-ReID-light, and DeepChange.Source: §1 · p. 1
  • The paper frames CC-ReID as a trade-off between inter-person discrimination and cross-clothing invariance, adaptively balancing dataset-dependent variance.Its stated contributions combine stochastic hard mining with multi-positive contrastive learning over hard-negative identities and cross-clothing positives.Source: §1 · p. 1

2. Related Work

Prior ReID work spans identification and metric-learning methods, backbone developments, and approaches that target clothing-invariant cues. MINER adds structured hard mining and multi-positive contrastive learning, while extending its use to cross-modal ReID.

Person Re-Identification

  • Earlier ReID methods use identification losses and triplet-style metric learning, while later work uses vision transformers and large-scale pretraining or self-supervision.MINER is described as backbone-agnostic and applicable to CNN-, ViT-, and Swin-based architectures.Source: §2 · p. 2

Clothes-Changing Re-Identification

  • Prior CC-ReID methods target clothes-agnostic cues through structural priors, semantic modeling, cloth-aware transformers, and identity or attribute disentanglement.The cited examples include body shape, silhouette, pose, parsing, CLIP-driven representations, and face/body fusion.Source: §2 · p. 2

Cross-Modal ReID

  • The paper also applies MINER to RGB-to-Depth ReID, using the same-clothes setting to demonstrate generality to a cross-modality gap.RGB-Depth maps RGB queries to a Depth gallery.Source: §2 · p. 2

Hard-Sample Mining in (CC-)ReID

  • Hard mining in ReID ranges from within-batch positives and negatives to global mining; CC-ReID approaches also generate or implicitly mine hard examples.Examples include semantic-guided hard-sample generation and cloth-aware augmentation or pose-diversified synthesis.Source: §2 · p. 2

Clothes-Changing Re-Identification

  • MINER samples visually similar identity centroids as hard negatives and dissimilar cross-camera clothing pairs as hard positives, producing batches with controlled intra- and inter-person difficulty.The soft sampling is compatible with standard contrastive training.Source: §2 · p. 2

Contrastive Learning for CC-ReID

  • MINER reformulates contrastive learning over multi-positive hard-mined batches organized across identities, cameras, and clothing sets, using an InfoNCE-style loss.This contrasts with prior CC-ReID work that often adapts CLIP-style supervision to emphasize biometric cues while mitigating clothing bias.Source: §2 · p. 2

3. Method

MINER combines stochastic hard mining with multi-positive contrastive learning, then adapts batch composition online using gradient-variance estimates to balance inter- and intra-person variation.

3.1. MINER Sampler

  • The sampler stochastically selects visually similar distinct identities, then samples cross-camera clothing pairs with low centroid similarity using a soft argmin.The selected identities are sampled without replacement, avoiding deterministic nearest-identity selection under noisy centroids.Source: §3.1 · p. 3

3.2. Contrastive Loss Function

  • A symmetric query-gallery cross-entropy uses block-diagonal positive targets to align cross-clothing matches and distinguish mined hard-negative identities.Each diagonal identity block contains multiple positive targets; off-diagonal blocks encode cross-identity negatives.Source: §3.2 · p. 4, Figure 1 · p. 4

3.3. Adaptive Batch Composition

  • Under a fixed image budget, MINER allocates identities and images per identity by minimizing combined between-identity and within-identity estimation risk.The closed-form optimum depends on the ratio of inter- to intra-person gradient variance and is updated online without extra forward passes or grid search.Source: §3.3 · pp. 4–5
  • Under memory constraints, PRCC mAP peaks at m=5 PIDs and n=17 images per person, whereas Celeb-L mAP peaks at m=21 PIDs and n=4 images per person, motivating dataset-adaptive allocation.Source: Figure 2 · p. 5, Figure 3 · p. 5
  • The adaptive allocation affects batch shape, not the contrastive objective, and relies on equal risk weighting, decoupled variance components, and projection-head gradients as a proxy.The authors note that the covariance assumptions are evaluated under conditional mining rather than IID sampling.Source: §3.3 · pp. 5–6

4. Balancing the Batch Variance

The best batch allocation depends on the dataset’s intra- and inter-person variance: PRCC and Celeb-L favor different identity-to-sample ratios. The adaptive controller recovers dataset-specific optima with lower regret than fixed allocations.

4. Balancing the Batch Variance

  • The same identity has greater intra-person variance in Celeb-L than PRCC, spanning clothing, pose, and scene.On PRCC, the allocation sweep peaks at ⟨5, 17⟩ and falls on either side of that point.Source: Figure 4 · p. 6, §4 · p. 6
  • Celeb-L peaks at ⟨21, 4⟩, while controlled-condition PRCC peaks at ⟨5, 17⟩, showing that dataset-specific variance shifts the preferred allocation.The PRCC peak reaches 69.7% mAP; the m-heavy ⟨29, 3⟩ and n-heavy ⟨2, 42⟩ allocations each reach 67.6%.Source: §4 · p. 6
  • 0.2 mAP worst-case regret and 0.10 mean regret let the controller outperform every fixed split, whose best worst-case and mean regrets are 0.6 and 0.33.The controller’s regrets at the PRCC, LTCC, and Celeb-L optima are 0.0, 0.1, and 0.2, respectively.Source: §4 · p. 6

5. Experiments

MINER improves CC-ReID retrieval across datasets and backbones, with state-of-the-art scores on four datasets. It also remains effective in same-clothes ReID and ranks on the podium across all three TVRID tracks.

5. Experiments

  • The evaluation trains a separate model per dataset using full dataset isolation and no re-ranking.MINER is finetuned from pretrained checkpoints for 20 epochs.Source: §5 · pp. 6–7
  • MINER achieves state-of-the-art scores on four CC-ReID datasets, while VC-Clothes and LaST remain competitive.The variance controller closely matches the validation-grid allocations on PRCC, Celeb-L, and LTCC without grid search.Source: §5 · p. 7
  • On PRCC, LTCC, and Celeb-L, MINER improves all evaluated backbones by 6.7 Rank@1 and 4.6 mAP points.Reported gains over vanilla pretrained EVA-CLIP and the ID-Loss baseline are (33.3, 32.5) and (6.5, 5.6) (∆R@1, ∆mAP), respectively.Source: §5 · p. 7
  • MINER reaches same-clothes ReID (Rank@1, mAP) scores of (100.0, 99.4) on PRCC and (93.8, 67.7) on LTCC.The evaluation compares against seven methods.Source: §5 · p. 7
  • Across TVRID’s RGB-RGB, Depth-Depth, and RGB-Depth tracks, MINER is the sole method to reach the podium among 148 teams and 732 submissions.Its respective (Rank@1, mAP) scores are (99.8, 99.9), (97.2, 98.3), and (91.5, 94.9).Source: §5 · p. 7

6. Discussion

The discussion argues that clothes-changing ReID requires explicit control of within-identity variation as well as between-identity separation. MINER combines stochastic hard mining with dataset-adaptive allocation, and its reported benefits extend across backbones and related ReID settings.

6. Discussion

  • Controlling within- and between-identity variance is a first-order factor for clothes-agnostic embeddings, rather than treating cross-outfit variation as a byproduct of metric learning.In CC-ReID, the dominant appearance shift is identity-preserving clothing change, and inter-person separation alone can leave embeddings brittle across outfits.Source: §6 · p. 7
  • Deterministic hardest-case mining reduces R@1 by 3.7/0.6 points on PRCC/Celeb-L relative to MINER; simultaneous hard-positive and hard-negative mining significantly improves R@1 (pvalue < 0.01).MINER samples softly over identity and outfit centroids to avoid instability from noisy hardness estimates.Source: §6 · p. 7, Table 3 · p. 8
  • MINER’s gains hold across CLIP-ReID, MADE, DIFFER, and EVA-CLIP, and extend to same-clothes and RGB-Depth ReID.The discussion relates modality variation in RGB-Depth ReID to clothing variation in CC-ReID.Source: §6 · p. 7
  • The optimal m:n allocation is dataset-dependent: controlled, homogeneous PRCC favors larger n, while in-the-wild Celeb-L favors larger m.Moving away from either dataset’s optimum reduces mAP; too few identities weakens inter-person pressure, while too few samples per identity weakens cross-outfit alignment.Source: §6 · p. 7
  • On a single A100 80GB GPU, the variance controller adds 0.7 ms per T steps to update batch composition.The runtime passage reports 54 seconds to embed PRCC, 13 ms to compute identity centroids, and 517 ms per SGD step.Source: §6 · p. 7
  • Table 3 reports sampler-ablation R@1 on PRCC/Celeb-L, where MINER uses both hard-positive and hard-negative mining; Table 2 reports Rank@1 and mAP across three CC-ReID datasets, with both metrics improved on all backbones and datasets.Source: Table 3 · p. 8, Table 2 · p. 8
  • MINER assumes vanilla-pretrained weights and clothing-ID labels, and requires at least two outfits per identity to expose intra-person variation.Training from scratch remains unexplored because early embeddings can yield noisy centroids and unstable mining; limited clothing diversity weakens hard-positive mining, as for some LTCC identities with one clothing ID.Source: §6 · p. 8

7. Conclusion

MINER addresses dataset-dependent clothing variation by balancing intra- and inter-person variance during training. The paper reports improved retrieval across CC-ReID benchmarks and generalization to conventional and RGB-Depth ReID.

7. Conclusion

  • Clothing-induced within-identity variation differs substantially across CC-ReID datasets.Source: §7 · p. 8
  • MINER is a stochastic mining framework with online closed-form batch adaptation that balances intra- and inter-person variance.Source: §7 · p. 8
  • MINER improves retrieval across CC-ReID benchmarks and generalizes to conventional and RGB-Depth ReID.Source: §7 · p. 8

Supplemental Material for MINER: Modeling Inter- and Intra-Person Variance for Reidentification

The supplement documents MINER’s reproduction settings and extended evaluations, then tests its sampler, Variance Controller, and contrastive objective. It reports dataset-dependent mining and allocation behavior alongside checks of the controller’s assumptions and objective formulation.

Supplemental analysis

  • Paired-seed ablations find hard-negative mining reliable across datasets, while hard-positive mining contributes most where intra-person variance is highest.The supplement also compares stochastic with deterministic hard mining and evaluates the objective and adaptive allocation independently of the sampler.Source: Supplemental Material for MINER:… · p. 12
  • The sampler selects which pairs enter a batch, whereas the Variance Controller allocates a fixed image budget between identities and samples per identity.Nine fixed allocations are compared; the controller has the lowest mean and worst-case regret and the best average rank.Source: Supplemental Material for MINER:… · p. 12
  • The closed-form allocation risk orders measured retrieval on PRCC and Celeb-L, while LTCC is a near-flat plateau with no meaningfully preferred allocation.The supplement also reports that equal weighting is immaterial, the omitted cross-term is small and sign-favorable, and a projection-head proxy recovers the full-gradient allocation within one unit per coordinate.Source: Supplemental Material for MINER:… · p. 12
  • The allocation objective is strictly convex along the budget frontier, and the gradient-level analysis relates MINER’s contrastive objective to supervised contrastive learning.Bipartite query–gallery normalization avoids the target-mass dilution attributed to an all-to-all formulation.Source: Supplemental Material for MINER:… · p. 12

A. Implementation Details

The supplement specifies MINER’s training and evaluation settings and reports results on clothes-changing, same-clothing, and cross-modal ReID. Its reported outcomes include a new LTCC-General state of the art, gains over EVA-CLIP baselines, and top-three TVRID placements in every track.

A. Implementation Details

  • The supplement specifies DIFFER and EVA2-CLIP-L/14 encoders with 1,024-D embeddings, 224 × 224 inputs, and MINER-specific batch and sampling settings.The default image budget ceiling is Bmax = 174; EVA-CLIP runs use 30 epochs at 1×10−6, and TVRID uses τ = 0.2.Source: A. Implementation Details · p. 12, Table 4 · p. 12

B. Experiments

  • Evaluation uses single-query retrieval without re-ranking or test-set statistics, reporting Rank@1 accuracy and mean Average Precision under each benchmark’s stated protocol.The protocols span cross-clothes camera splits, clothes-changing splits, and a large-scale spatio-temporal protocol.Source: B. Experiments · pp. 12–13

B.1. Clothes-Changing ReID

  • The clothes-changing evaluation covers six datasets under their cross-clothing protocols, including PRCC, LTCC, Celeb-ReID-Light, DeepChange, VC-Clothes, and LaST.The corresponding results are reported in Table 1 of the main paper.Source: §B.1 · p. 13

B.2. Traditional ReID

  • On LTCC-General, MINER reaches 93.8/67.7 versus the next-best 85.0/52.8 in R@1/mAP, while PRCC-SC R@1 is saturated at 100.0 across seven methods.The supplement characterizes MINER’s PRCC-SC mAP of 99.4 as competitive but not above the top entries.Source: §B.2 · p. 13, Table 5 · p. 13

B.3.1. Backbone Gains

  • Adding MINER as a finetuning stage consistently improves reproduced backbone scores, with matched base and MINER runs isolating the gain under identical settings.The supplement presents this as evidence of MINER’s value as an architecture-agnostic finetuning step.Source: §B.3.1 · p. 13

B.3.2. Baseline

  • MINER adds mAP gains of 46.1, 29.3, and 22.0 over vanilla EVA-CLIP on PRCC, LTCC, and Celeb-L, and 7.0, 4.0, and 5.8 over the ID-loss baseline.The comparison covers EVA-CLIP off the shelf, ID-loss-finetuned EVA-CLIP, and ReID-pretrained DIFFER under the clothes-changing protocol.Source: §B.3.2 · p. 13, Table 6 · p. 13

B.4. TVRID 2026 Competition

  • MINER is the only method in the TVRID competition’s top three across all three tracks, and the RGB-Depth podium score is 94.9 mAP versus a 40.7 track average.The reported podium-to-average gap indicates the setting is far from saturated; the authors note that top-view sensors hide faces.Source: §B.4 · pp. 13–14, Table 7 · p. 14

B.5. MINER’s Sampler: Ablation Study

The sampler ablation separates the contributions of hard-negative and hard-positive mining and compares stochastic hard mining with deterministic selection. The full sampler improves Rank@1 on both datasets, with the component contributions varying by dataset.

Sampler ablation

  • The full sampler yields the largest significant Rank@1 gains on PRCC (+6.50, pvalue = 0.0007) and Celeb-L (+2.17, pvalue = 0.0059).Hard-negative mining significantly improves over random mining on both datasets; hard-positive mining is not significant on PRCC and only marginal on Celeb-L, where its point estimate is larger.Source: §B.5 · p. 14
  • MINER’s stochastic hard mining exceeds deterministic hard mining by 3.70 R@1 points on PRCC and 0.56 on Celeb-L.Deterministic hard mining still outperforms random mining, and the smaller stochasticity gap on Celeb-L shows that this benefit varies by dataset.Source: §B.5 · p. 14

B.6. Multi-Positive Objective and Variance Controller: Ablation Study

The ablation finds that MINER’s objective and adaptive allocation contribute beyond hard-example mining. The accompanying cross-modal figure depicts how the contrastive loss represents inter-person and intra-person variance.

B.6. Multi-Positive Objective and Variance Controller: Ablation Study

  • With hard mining disabled, the multi-positive objective and Variance Controller improve R@1 over matched ID-loss baselines by +2.0 points on PRCC and +1.6 on Celeb-L.The PRCC comparison is 65.1% to 67.1%, and the Celeb-L comparison is 72.7% to 74.3%.Source: §B.6 · pp. 14–15
  • Figure 5 depicts a cross-modal bipartite-batch contrastive loss, with hard-negative identities and hard-positive cross-modal pairs encoding inter-person variance in white and intra-person variance in blue.The block-diagonal label matrix is shown for the TVRID RGB-Depth data.Source: Figure 5 · p. 15

B.7. MINER’s Variance Controller: Ablation Study

The Variance Controller improves on fixed batch allocations across the measured grid, while the accompanying analyses test whether its risk model and gradient-based estimates support its allocation choices.

B.7. MINER’s Variance Controller: Ablation Study

  • The controller’s worst-case regret is 0.2 mAP and mean regret is 0.10, versus 0.6 and 0.33 for the best fixed allocation on each criterion.It Pareto-dominates all nine fixed allocations on both regret measures.Source: §B.7 · pp. 15–16, Table 10 · p. 17
  • The controller achieves a best mean rank of 3.33, ahead of the next-best fixed split at 4.50.The three aggregate criteria each favor a different fixed allocation, while the controller leads on all three.Source: §B.7 · p. 16
  • With only three datasets, the omnibus rank test is underpowered, so the regret gaps are reported descriptively rather than as inferential claims.The critical difference of approximately 6.9 ranks exceeds the observed rank range of 5.5.Source: §B.7 · p. 16

B.8. Variance Controller Assumption Validation

  • Changing λ across 0.2, 0.5, and 0.8 shifts the Celeb-L allocation substantially, while mAP and R@1 stay within seed noise.The equal-weighting choice therefore does not affect retrieval in this test.Source: §B.8 · p. 16, Figure 7 · p. 18
  • The additive risk omits a small negative cross-term, making it a conservative bound while preserving the selected allocation.The additive allocation matches the grid optimum and full-parameter solution.Source: §B.8 · p. 17, Table 11 · p. 18
  • The projection-head estimator recovers the full-parameter allocation exactly on PRCC and within one unit per coordinate on LTCC.The controller uses projection-head gradients rather than full-parameter gradients.Source: §B.8 · pp. 17–18, Table 11 · p. 18

C.1. Per-Query Gradient and the Two-Signal Split

  • Positive within-identity pairs pull together cross-clothing matches, while negative pairs push apart cross-identity look-alikes, forming invariance and discrimination gradient signals.The gradient is split into within-block positive-pair and off-block negative-pair contributions.Source: §C.1 · p. 18

C.2. Convexity of the Allocation Objective

  • Under the non-degenerate positive-trace case, the allocation objective is strictly convex along the saturated budget frontier, yielding a unique minimizer projected to feasible integer allocations.The budget constraint is saturated at 2mn = Bmax.Source: §C.2 · pp. 18–19

C.3. Loss Function Analysis

The loss analysis argues that bipartite normalization better matches retrieval than all-to-all normalization: it preserves target mass on query–gallery comparisons and avoids adding within-set pairs that dilute or duplicate that signal.

C.3. Loss Function Analysis

  • MINER’s sampler concentrates hard negatives that produce disproportionately large gradients, while its block-diagonal positives distribute gradient mass across cross-clothing matches.The analysis links these structures to discriminative pressure and clothing-invariant identity clusters.Source: §C.3 · p. 19
  • At the PRCC and Celeb-L optima, cross-set positives receive 17/33 ≈ 0.52 and 4/7 ≈ 0.57 of positive target mass, respectively, under all-to-all normalization.Bipartite normalization assigns total positive-target mass of 1 to cross-set positives.Source: §C.3 · p. 20
  • When within-set candidates are nonnegligible, they shrink cross-set probabilities by a common factor, uniformly attenuating cross-set gradients as well as diluting target mass.The probability shrinkage acts like a smaller effective step size and compounds the target-mass effect.Source: §C.3 · p. 20
  • Under ℓ2-normalization and strong initialization, same-identity query pairs become nearly collinear, making their positive-gradient contribution negligible while their logits drain cross-set gradient budget.The update direction is therefore dominated by cross-set terms.Source: §C.3 · pp. 20–21
  • Within-set negatives duplicate identity-separation pressure already supplied by cross-set query–gallery negatives, which the sampler selects to be maximally confusable.Their utility is dataset- and regime-dependent, while they incur the dilution quantified above.Source: §C.3 · p. 21
Loading 2610.28941v2…