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

Athena Psalta; Vasileios Tsironis; Konstantinos Karantzalos

cs.CVOct 25, 2026

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

Authors: Athena Psalta, Vasileios Tsironis, Konstantinos Karantzalos

Organizations: National Technical University of Athens, Iroon Polytechneiou 9, Athens 15780, Greece

Abstract

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.

Figures & tables

Figure 1: PBR-ReID Framework. Input images are passed through a DiNOv3 ViT backbone to produce normalized embeddings, jointly optimized with a learnable prototype matrix (one vector per identity) that anchors each class in the embedding space. The staged training progression (right) illustrates how embedding geometry evolves from random initialization through prototype warmup to joint fine-tuning, with features forming compact intra-class clusters separated by enforced inter-class margins. Prototypes act only during training and are discarded at inference.

Figure 2: Boundary Repulsion MixUp (BRM). BRM operates in four stages. First, pairwise prototype cosine similarities are computed for all identities in the batch. Second, confusable pairs (those exceeding similarity threshold $\tau$ ) are identified. Third, for each confusable pair, a boundary-proxy prototype is constructed by linearly interpolating the two class prototypes with a Beta-sampled coefficient $\lambda$ , which also mixes the corresponding features. Fourth, the BRM loss combines hard-target attraction terms, which pull the mixed feature toward both participating prototypes, with repulsion terms, which penalize original features lying within margin $m_{\mathrm{BRM}}$ of the boundary-proxy prototype.

Model	Backbone	#Params	Market-1501		DukeMTMC		MSMT17		CUHK03-NP		IUST
Model	Backbone	#Params	mAP	Rank-1	mAP	Rank-1	mAP	Rank-1	mAP	Rank-1	mAP	Rank-1
AdaSP [ 69 ]	ResNet-50	25M	89.8	95.5	83.0	91.7	67.1	85.5	80.1	82.0	-	-
ProNet++ [ 50 ]	ResNet-50	25M	90.2	96.0	-	-	65.5	85.4	82.7	85.2	-	-
CA-Jaccard [ 4 ]	ResNet-50	25M	86.1	94.4	-	-	44.3	75.1	-	-	-	-
TMGF [ 21 ]	ViT-S/16	21M	91.9	96.3	83.1	92.3	70.3	88.2	-	-	-	-
TransReID [ 13 ]	ViT-B/16	86M	89.5	95.2	82.6	90.7	69.4	86.2	72.9	75.1	-	-

Table 1: Person ReID Results. Simple ViT-B/16 denotes an ImageNet pre-training; ∗ pre-trained on LUPerson [ 10 ] ; † CLIP ViT-L/14 used as a distillation teacher with a ResNet-50 student; ‡ combines DINOv2 ViT-B/14 and CLIP ViT-B/16.

Model	VeRi-776		VehicleID
Model	mAP	Rank-1	Rank-1	Rank-5
CFA-Net [ 48 ]	80.7	96.9	-	-
APC [ 52 ]	85.1	97.9	-	-
VehicleNet [ 65 ]	83.4	96.8	79.5	92.0
TANet [ 15 ]	83.6	96.8	78.2	92.6
HCI-Net [ 42 ]	83.8	96.6	76.4	91.2

Table 2: Vehicle ReID Results.

	Market-1501		DukeMTMC
Backbone	mAP	Rank-1	mAP	Rank-1
ViT-B/16 (ImageNet)	92.2	96.3	84.1	91.8
CLIP ViT-B/16	93.4	96.6	85.4	92.7
DINOv2 ViT-B/14	93.7	96.7	85.9	92.9
DINOv3 ViT-B/16	94.0	97.1	86.1	93.1

Table 3: Effect of the backbone encoder. We evaluate PBR-ReID with different pre-trained ViT-B/16 backbones on Market1501 [ 63 ] and DukeMTMC [ 64 ] . All variants use identical training and inference settings.

	Market-1501		DukeMTMC
Training Objective	mAP	Rank-1	mAP	Rank-1
ArcFace [ 6 ]	90.2	94.9	82.0	90.7
CosFace [ 47 ]	89.6	95.8	82.6	92.4
CircleLoss [ 43 ]	91.0	96.0	83.9	92.4
VPL [ 7 ]	90.0	95.0	81.9	90.6
PBR-ReID (Ours)	94.0	97.1	86.1	93.1

Table 4: Comparison with prototype-based objectives. All variants share the same DINOv3 ViT-B/16 backbone and training protocol, only the prototype-based objective differs.

$\mathcal{L}_{\mathrm{bht}}$	$\mathcal{L}_{\mathrm{fpt}}$	$\mathcal{L}_{\mathrm{cen}}$	$\mathcal{L}_{\mathrm{norm}}$	MixUp	BRM	Market-1501		DukeMTMC
$\mathcal{L}_{\mathrm{bht}}$	$\mathcal{L}_{\mathrm{fpt}}$	$\mathcal{L}_{\mathrm{cen}}$	$\mathcal{L}_{\mathrm{norm}}$	MixUp	BRM	mAP	Rank-1	mAP	Rank-1
✓						88.9	94.5	82.1	90.8
✓			✓			90.4	95.2	83.2	91.3
✓	✓					86.3	93.0	79.8	88.7
✓	✓	✓				87.5	93.7	80.7	89.9
✓	✓	✓	✓			92.4	96.3	84.6	92.4

Table 5: Effect of loss terms & BRM. The top block uses instance-level supervision only (no prototypes); the bottom block introduces them. NG: neighbor-guided MixUp.

Panel of Figure 3 — Figure 3: Ablation studies. Hyperparameter ablations (left) and KDEs of inter-class feature distances on Market-1501 (right), warmup epochs experiments (center)

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

		Market-1501		DukeMTMC
Backbone	#Params	mAP	Rank-1	mAP	Rank-1
ViT-B/16	86M	94.0	97.1	86.1	93.1
ViT-L/16	300M	94.8	97.2	89.7	94.6
ViT-H+/16	840M	95.3	97.4	90.8	95.1

Appendix

Table 6: Performance scaling with backbone capacity. PBR-ReID with DINOv3 ViT backbones experiments with fixed training and inference settings.

Dual LR	EMA Decay	Market-1501		DukeMTMC
Dual LR	EMA Decay	mAP	Rank-1	mAP	Rank-1
✓	0.999	94.0	97.1	86.1	93.1
✓	0.90	92.1	95.9	84.4	91.5
✓	0 (No EMA)	81.7	90.8	75.0	85.2
	0.999	91.4	95.8	83.9	91.9

Appendix

Table 7: Ablation of training stabilization components. Effect of the dual learning-rate schedule and the EMA prototype update during joint fine-tuning. The top row is the full PBR-ReID configuration.

Method	CUHK03-D		CUHK03-L		VehicleID
					Small		Medium		Large
	mAP	Rank-1	mAP	Rank-1	Rank-1	Rank-5	Rank-1	Rank-5	Rank-1	Rank-5
CLIP-ReID [ 23 ]	-	-	-	-	85.3	97.6	81.0	95.0	78.1	92.7
AdaSP [ 69 ]	80.1	82.0	82.4	84.6	89.3	98.4	85.6	96.7	83.0	94.9
PBR-ReID (Ours)	88.4	89.8	89.7	90.6	94.0	97.8	92.1	96.6	90.2	96.0

Appendix

Table 8: Additional ReID results on CUHK03 [ 67 ] and VehicleID [ 26 ] benchmarks.

	Query	R-1				R-5					R-10
No BRM	ID: 66	ID: 65	ID: 65	ID: 66	ID: 65	ID: 65	ID: 65	ID: 65	ID: 65	ID: 65	ID: 65
BRM	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66	ID: 66

No BRM	ID: 615	ID: 585	ID: 585	ID: 585	ID: 615	ID: 585	ID: 615	ID: 585	ID: 585	ID: 585	ID: 585
BRM	ID: 615	ID: 615	ID: 615	ID: 615	ID: 615	ID: 615	ID: 615	ID: 615	ID: 585	ID: 615	ID: 615

Appendix

Figure 4: Top-10 retrieval results for 3 VeRi-776 queries (PIDs 66, 615, 38), comparing the baseline without BRM against the BRM-trained model. Green = correct match; red = incorrect match; blue = query image.

	Query	R-1				R-5					R-10
No BRM	ID: 702	ID: 2250	ID: 702	ID: 2250	ID: 2250	ID: 2250	ID: 2250	ID: 691	ID: 2250	ID: 2250	ID: 691
BRM	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702	ID: 702

No BRM	ID: 803	ID: 396	ID: 820	ID: 396	ID: 396	ID: 396	ID: 396	ID: 396	ID: 1931	ID: 1862	ID: 820
BRM	ID: 803	ID: 803	ID: 803	ID: 803	ID: 803	ID: 803	ID: 803	ID: 803	ID: 831	ID: 803	ID: 803

Appendix

Figure 5: Top-10 retrieval results for 3 MSMT17 queries (PIDs 702, 803, 1874), comparing the baseline without BRM against the BRM-trained model. Green = correct match; red = incorrect match; blue = query image.

	Query	R-1				R-5					R-10
No BRM	ID: 2046	ID: 2216	ID: 2216	ID: 2216	ID: 2216	ID: 2216	ID: 2073	ID: 61	ID: 1834	ID: 1834	ID: 2216
BRM	ID: 2046	ID: 2046	ID: 2046	ID: 2046	ID: 2046	ID: 2046	ID: 2046	ID: 2073	ID: 2229	ID: 2046	ID: 2046

No BRM	ID: 2055	ID: 1641	ID: 1641	ID: 2287	ID: 2287	ID: 2287	ID: 2055	ID: 2055	ID: 1641	ID: 1641	ID: 2287
BRM	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055	ID: 2055

Appendix

Figure 6: Top-10 retrieval results for 3 MSMT17 queries (PIDs 2046, 2055, 2635), comparing the baseline without BRM against the BRM-trained model. Green = correct match; red = incorrect match; blue = query image.

Person Re-Identification

Person re-identification matches the same individual across camera views, times, clothing, and sensing modalities despite changes in appearance.

See all 92 papers

DiffReID: Discriminative Diffusion Model for Object Re-Identification
Sep 29, 2026Yingquan Wang, Pingping Zhang, Dong Wang +1Representation Learning Query Re-Identification
School of Information and Communication Engineering, Dalian University of Technology · School of Future Technology, Dalian University of Technology
Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification
Apr 21, 2026Quan Zhang, Jingze Wu, Jialong Wang +3Person Re-Identification
Sun Yat-sen University · Alibaba Cloud Computing
Prototype Matters: Modality-unified Prototype Self-distillation for Unsupervised Visible-infrared Person Re-identification
Sep 10, 2026Menglin Wang, Xiaojin GongContrastive Learning Cross-Modal Alignment
Nanjing Normal University Nanjing, China · Zhejiang University Hangzhou, China

Sep 29, 2026J/K move · Enter open · S save

DiffReID: Discriminative Diffusion Model for Object Re-Identification

Yingquan Wang, Pingping Zhang, Dong Wang, Huchuan Lu

School of Information and Communication Engineering, Dalian University of Technology · School of Future Technology, Dalian University of Technology

Representation Learning Query Re-Identification Diffusion Models

Open paper AskPDF arXiv

As a fundamental image processing task, object Re-Identification (ReID) aims to retrieve objects across non-overlapping cameras. Recently, with the development of deep learning, significant advancements have been made in object ReID. However, most existing methods suffer from generalization due to the limited size and diversity of ReID datasets. Meanwhile, current models tend to focus on extracting semantic patterns rather than learning identity-aware feature distributions. To address these issues, we propose a novel feature learning framework named DiffReID for object ReID. It leverages a discriminative diffusion model to gradually learn identity-aware distributions and generate identity-invariant features. More specifically, with the Contrastive Language-Image Pre-training (CLIP) model, we first obtain identity-aware text features by prompt tuning. Then, we propose a Vision-guided Noise Generator (VNG) to initialize probabilistic noises and gradually corrupt identity-aware text features. Afterwards, we take visual features as conditions and propose a Light Weight Denoiser (LWD) to denoise the corrupted text features step-by-step for identity-aware distribution learning. To obtain discriminative features, we further generate identity-invariant guided features from randomly sampling visual-guided noises. Finally, we propose a Mutual Enhancement Constraint (MEC) to facilitate mutual learning between visual features and guided features to enhance the representation robustness and discrimination. Extensive experiments on five object ReID benchmarks demonstrate that our method shows better results than most state-of-the-art methods. The source code is available at https://github.com/AWangYQ/DiffReID.

Abstract

Person Re-Identification

Explore similar work

DiffReID: Discriminative Diffusion Model for Object Re-Identification

Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification

Prototype Matters: Modality-unified Prototype Self-distillation for Unsupervised Visible-infrared Person Re-identification