PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Discriminative Prototypes via Boundary Repulsion MixUp for Object Re-Identification

Athena Psalta, Vasileios Tsironis, Konstantinos Karantzalos

arXiv:2610.34305v1cs.CV

TL;DR

ReID must remain robust to visual variation while separating identities whose embeddings are confusable, but pairwise losses shape those boundaries only implicitly. PBR-ReID learns identity prototypes and uses their similarities to select boundary proxies for repulsion; with a DINOv3 ViT-B/16 backbone, it reports strong results across competitive benchmarks.

  • Problem

    ReID must distinguish visually similar identities under viewpoint, illumination, and occlusion variation, while the inter-class geometry of confusable identities remains only implicit in pairwise losses.

    Source: §1 · p. 1

  • Method

    PBR-ReID jointly learns identity prototypes with a DINOv3 backbone and uses prototype similarities to select confusable pairs whose interpolated boundary proxies repel encroaching features.

    Source: §1 · p. 2

  • Results

    PBR-ReID with a DINOv3 ViT-B/16 backbone reports strong results across competitive benchmarks.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The prototypes shape feature geometry during training and are discarded at inference, where retrieval relies on backbone features alone.

    Source: §1 · p. 2

Abstract

from arXiv · show

Object re-identification requires embeddings that are robust to intra-class variation in viewpoint, illumination, and occlusion, yet preserve sharp separation between visually similar identities. The hardest errors arise precisely where confusable identities meet, a region that pairwise losses shape only implicitly. We present PBR-ReID, a prototype-based framework that makes this inter-class geometry an explicit learning target. Each identity is represented by a learnable prototype jointly optimized with a DiNOv3 backbone and the resulting prototype similarity matrix reveals, at every training step, which identities are most confusable. On this signal we build Boundary Repulsion MixUp (BRM), a geometry-aware regularizer that converts standard MixUp's weakness into a discriminative signal. Rather than interpolating samples across confusable identities, which blurs the boundaries that matter most, BRM constructs an explicit boundary proxy between each confusable pair and repels their features away from it, converting boundary ambiguity into a discriminative signal. We adopt a two-stage training protocol consisting of a prototype warmup phase followed by EMA-stabilized joint fine-tuning. Extensive experiments on person and vehicle ReID benchmarks demonstrate strong performance, with thorough ablations confirming each component's contribution. Code is available at https://github.com/psaltaath/pbr-reid.

1. Introduction

Object Re-Identification must handle viewpoint, illumination, and occlusion while distinguishing visually similar identities. PBR-ReID makes learnable identity prototypes and confusable-class boundaries explicit targets during training.

  • ReID must match identities across non-overlapping cameras while remaining robust to viewpoint, illumination, and occlusion and separating visually similar identities.Source: §1 · p. 1
  • Prior prototype approaches leave inter-class geometry uncontrolled during training, whether prototypes are static centroids or gallery-dependent inference constructs.Centroids are confined to observed-feature convex hulls and can be brittle under multimodal intra-class distributions.Source: §1 · p. 1
  • Confusable identities concentrate near shared decision boundaries, where geometry-agnostic MixUp can reinforce rather than resolve ambiguity.Uniform interpolation can place synthetic samples between highly confusable identities.Source: §1 · pp. 1–2
  • PBR-ReID jointly optimizes a learnable prototype per identity with a DINOv3 backbone, using prototype similarities to expose pairwise identity relationships.Prototypes shape training-time feature geometry and are discarded at inference.Source: §1 · p. 2
  • Boundary Repulsion MixUp selects the most confusable pairs from prototype similarities, interpolates boundary proxies between their prototypes, and penalizes features encroaching on them.Training uses prototype warmup followed by EMA-stabilized joint fine-tuning.Source: §1 · p. 2

2. Related Work

Prior ReID work commonly treats the backbone as the main source of discriminability, while class structure emerges implicitly from pairwise losses. PBR-ReID instead uses prototype relationships to target confusable identities and their embedding boundaries.

  • ReID backbones span vision-language alignment, self-supervised or prompt-driven pretraining, and attention or part-based decomposition, while class structure often emerges implicitly from pairwise losses.Source: §2 · p. 2
  • Earlier prototype methods use memory anchors, contrastive targets, classifier weights, or gallery-derived representatives, but do not make prototype pairwise structure a training signal.Inference-time prototype methods can tie the model to the gallery.Source: §2 · p. 2
  • Common metric and augmentation regularizers leave confusable identities near embedding boundaries unaddressed; BRM targets those pairs using prototype similarities and boundary proxies.The passage contrasts BRM with triplet objectives and uniformly interpolated pairs.Source: §2 · p. 2

3. Methodology

PBR-ReID jointly learns class prototypes and normalized features, then uses BRM to target confusable boundaries; staged optimization ends in feature-only retrieval.

3.1. Feature Representation

  • Learnable identity prototypes anchor normalized DiNOv3 embeddings and define a batch similarity matrix that represents inter-class confusion.The matrix is recomputed for sampled identities; its construction scales with batch size rather than the total number of identities.Source: §3.1 · pp. 2–3

3.2.1. Training Objective

  • A five-term objective combines feature and prototype triplets, center loss, norm regularization, and BRM to balance class compactness with inter-class separation.Class-balanced batches provide positive pairs, negative diversity, and an online estimate of prototype similarities; batch-hard feature triplet and BRM activate only during joint fine-tuning.Source: §3.2.1 · p. 3

3.2.2. Boundary Repulsion MixUp (BRM)

  • BRM identifies confusable identity pairs, mixes their features and prototypes, attracts mixed features to both prototypes, and repels original features from the boundary proxy.Repulsion also drives the paired prototypes apart, while adaptive separation is intended to stop once it clears intra-class spread.Source: §3.2.2 · pp. 4–5

3.2.3. Training Strategy & Inference

  • Training first warms up prototypes with a frozen backbone, then jointly fine-tunes with EMA prototype updates; inference discards prototypes and retrieves by normalized embeddings.BRM and batch-hard feature triplet are disabled during warmup and activated during joint fine-tuning.Source: §3.2.3 · p. 5

4. Experiments

PBR-ReID is evaluated on person and vehicle benchmarks, with strong comparisons across both domains. Ablations examine backbone and loss choices, BRM’s effect on confusable identities, and training stability and sensitivity.

4.1. Experimental settings

  • Evaluation spans five person ReID benchmarks and two vehicle ReID datasets; performance is assessed using Cumulative Matching Characteristic (CMC) curves and mean Average Precision (mAP).The person benchmarks are Market-1501, DukeMTMC [64], MSMT17 [53], CUHK03 [67], and IUST [32]; the vehicle datasets are VeRi-776 and VehicleID.Source: §4.1 · pp. 5–6

4.2. State-of-the-art comparison

  • PBR-ReID outperforms compared methods on four person benchmarks and ranks first on two vehicle benchmarks, including 88.0 mAP and 98.1 rank-1 on VeRi-776.On MSMT17, it leads HPL by +0.7% mAP and trails DRFormer by 0.1% rank-1; on VehicleID Large, it reaches 90.2 rank-1.Source: §4.2 · pp. 6–7, Table 1 · p. 6, Table 2 · p. 7

4.3. Ablation studies

  • Across backbone variants, PBR-ReID surpasses the strongest same-family competitor, while self-supervised backbones outperform CLIP and supervised variants.Against prototype-based objectives on the same DINOv3 backbone, it improves mAP over the strongest baseline by 3.0% on Market-1501 and 3.5% on DukeMTMC.Source: §4.3 · p. 7, Table 3 · p. 7, Table 4 · p. 7
  • Unit normalization lifts mAP by +4.9% on Market-1501 and +3.9% on DukeMTMC, enabling the prototype configuration to reach 92.4% and 84.6% mAP, respectively.Detaching prototypes from center loss lowers mAP by 1–1.5% across the two datasets.Source: §4.3 · p. 7, Table 5 · p. 8
  • BRM improves mAP over the prototype baseline by +1.6% on Market-1501 and +1.5% on DukeMTMC; neighbor-guided MixUp instead falls below the no-MixUp baseline.Because both variants select the same confusable pairs, the comparison isolates repulsion versus interpolation.Source: §4.3 · p. 8, Table 5 · p. 8
  • Confusable-pair mean distance rises from 0.2393 to 0.2514; without warmup, mAP collapses to 51.4% on Market-1501 and 50.9% on DukeMTMC.Performance peaks at 10 warmup epochs and remains stable for longer schedules; EMA removal costs 12.3% and 11.1% mAP, while BRM hyperparameter sweeps vary by at most 1.1 points.Source: §4.3 · p. 8, Figure 3 · p. 8

5. Conclusions

PBR-ReID makes inter-class geometry an explicit learning target through learnable identity prototypes and Boundary Repulsion MixUp. The authors report that its gains are not attributable to the backbone and suggest extending the approach to unsupervised ReID.

5. Conclusions

  • PBR-ReID uses a learnable prototype per identity to make inter-class geometry an explicit learning target for object re-identification.The prototypes and BRM shape training; retrieval uses backbone features alone.Source: §5 · p. 8
  • The authors report that ablations attribute gains to the method rather than the backbone and find performance robust to its hyperparameters.The conclusion connects both findings to the reported PBR-ReID results.Source: §5 · pp. 8–9
  • A proposed next direction is unsupervised ReID using prototypes derived from cluster pseudo-labels instead of ground-truth identities.Source: §5 · p. 9

A. Additional Experiments

Additional experiments report results on CUHK03 and VehicleID, examine backbone scaling and inference cost, and ablate the joint fine-tuning stabilization mechanisms.

Complementary ReID results

  • CUHK03-L reaches 89.7 mAP and 90.6 Rank-1, versus 88.4 mAP and 89.8 Rank-1 on CUHK03 detected; VehicleID Rank-1 leads across splits, from 94.0 Small to 90.2 Large.Source: A. Additional Experiments · p. 12

Performance scaling & Efficiency

  • Inference uses one backbone pass followed by nearest-neighbour matching; accuracy improves with backbone capacity, while ViT-B/16 is competitive at a fraction of the parameters.The prototype matrix and BRM are discarded before retrieval.Source: A. Additional Experiments · p. 12, Table 6 · p. 12

Experiments on training stabilization

  • Disabling EMA lowers mAP to 81.7 on Market-1501 and 75.0 on DukeMTMC, while a single shared learning rate yields 91.4% and 83.9%, respectively.The authors attribute the EMA loss to noisy prototypes producing unreliable boundary proxies; a 0.90 decay trails 0.999 by 1.9 and 1.7 mAP.Source: A. Additional Experiments · p. 12, Table 7 · p. 12

B. Visualizations

Qualitative top-10 retrieval comparisons show BRM changing results in deliberately difficult, confusable cases across person and vehicle ReID. The examples are selected to expose boundary cases and are not representative of typical retrieval.

B. Visualizations

  • The queries are deliberately drawn from the hardest confusable cases to expose BRM's effect, so they are not representative of typical retrieval.For each query, the baseline list is shown above the BRM-trained model's list.Source: B. Visualizations · p. 12
  • On MSMT17, BRM promotes correct matches and suppresses near-boundary confusions from pedestrians sharing clothing style and color, including under strong appearance shifts.Figures 5 and 6 show top-10 lists; green marks correct matches, red incorrect matches, and blue the query.Source: B. Visualizations · p. 12, Figure 5 · p. 14, Figure 6 · p. 15
  • On VeRi-776, BRM recovers correct matches where baseline top ranks feature wrong-identity vehicles of the same model and color.Figure 4 compares top-10 retrievals; green marks correct matches, red incorrect matches, and blue the query.Source: B. Visualizations · p. 12, Figure 4 · p. 13
Loading 2610.34305v1…