PaperlayerView exact revision on arXiv

Source-linked AI summary

FM-ReID: Selective Competitive Token Routing for Object Re-Identification

Zhiqi Li, Xiaowei Zhou, Zeyuan Sun, Feng Gao, Junyu Dong

arXiv:2609.36560v1cs.CVcs.AI

TL;DR

ReID must distinguish identities with localized cues that can be hard to capture in a holistic descriptor, while fixed partitions and independent query normalization have limitations. FM-ReID routes dense DINOv3 tokens through competing queries and jointly trains multi-query descriptors with a holistic representation; evaluations across animal, person, and vehicle benchmarks support competitive token routing as an augmentation to holistic representations.

  • Problem

    Identities can share similar global appearances, while their distinguishing cues are localized and viewpoint-dependent; fixed partitions and independently normalized queries also have limitations for local representation learning.

    Source: §1 · p. 2

  • Method

    FM-ReID uses competing mining and residual queries with above-prior token selection, then jointly trains multi-query descriptors with a holistic representation for retrieval.

    Source: §1 · p. 2

  • Results

    Evaluations on animal, person, and vehicle ReID benchmarks show that competitive token routing can effectively augment holistic foundation-model representations.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Separate supervised experiments on animal, person, and vehicle benchmarks support the architecture’s cross-category applicability.

    Source: §5 · p. 8

    Transfer to unseen object categories remains unevaluated.

    Source: §5 · p. 9

Abstract

from arXiv · show

Object re-identification (ReID) faces a recurring challenge: different identities can share highly similar global appearances, while the cues that distinguish them are localized, heterogeneous, and visible only under particular viewpoints. This challenge arises in animal ReID through markings, contours, and scars, in person ReID through subtle clothing and accessory cues, and in vehicle ReID through localized appearance details. Although visual foundation models encode such information in dense tokens, a single holistic descriptor can obscure discriminative local signals. We propose FM-ReID, an end-to-end framework that formulates local representation learning as selective competitive token routing. Its Competitive Fine-grained Mining module uses multiple mining queries and a residual query to compete for dense DINOv3 tokens. Above-prior selection retains tokens preferentially allocated to each mining query, while the residual slot receives tokens excluded from the retrieval descriptors. The resulting multi-query descriptors are jointly trained with a holistic representation for retrieval, without fixed spatial partitions or equal-area constraints. FM-ReID achieves strong results on animal, person, and vehicle ReID benchmarks, supporting competitive token routing as an effective way to augment holistic foundation-model representations.

1 Introduction

ReID must distinguish identities whose localized cues vary in visibility and extent, while fixed partitions and independently normalized queries can limit complementary local representations. FM-ReID routes dense DINOv3 tokens competitively and jointly trains the resulting fine-grained descriptors with a holistic representation.

1 Introduction

  • Different identities can share similar global appearances, while distinguishing cues are localized, heterogeneous, and visible only from particular viewpoints.Such cues include animal markings, contours, or scars, subtle clothing and accessories, and localized vehicle details.Source: §1 · p. 2
  • Fixed partitions and rearrangement rules rely on approximate spatial alignment, while independently normalized queries may repeatedly attend to the same salient evidence.Dense foundation-model tokens encode spatial and semantic structure, but an allocation mechanism must distinguish query-specific evidence from ambiguous or uninformative tokens.Source: §1 · p. 2
  • FM-ReID augments a holistic representation with fine-grained features selectively mined from dense foundation-model tokens, without part annotations.It jointly trains holistic and multi-query descriptors for retrieval, with evaluations spanning animal, person, and vehicle ReID benchmarks.Source: §1 · p. 2
  • CFM makes mining queries and a residual query compete for dense tokens; above-prior selection permits unclaimed evidence and query supports of unequal size.The residual slot receives evidence excluded from retrieval descriptors, and the resulting multi-query descriptors are jointly trained with the holistic representation.Source: §1 · p. 2

2 Related Work

Related work explores local representations, query-based region discovery, and foundation-model features for ReID. FM-ReID draws on these directions while using selective competitive aggregation over dense tokens.

2.1 Fine-Grained Representation Learning for Object ReID

  • Prior methods use fixed stripes or granularities, token rearrangement, learned part prototypes, and anatomical tokens; FM-ReID instead learns selective aggregation without prescribed regions or equal-area assignments.Fixed partitions can become semantically inconsistent under pose changes, occlusion, or non-human geometry.Source: §2.1 · pp. 2–3

2.2 Query-Based Discovery and Competitive Assignment

  • Prior work explores latent prototypes, competitive assignment, balanced patch-to-part allocation, and separate identity-relevant queries; CFM combines competition with above-prior selection over dense tokens.Slot Attention provides an allocation principle that CFM adapts for discriminative retrieval, with a residual query and unequal query-specific support.Source: §2.2 · p. 3

2.3 Visual Foundation Models for ReID

  • Vision-language ReID emphasizes cross-modal knowledge or temporal modeling, while DINO-family models provide category-agnostic dense features that FM-ReID selectively aggregates without text prompts or secondary in-domain pre-training.FM-ReID evaluates the empirical contribution of this aggregation using the same DINOv3 backbone.Source: §2.3 · p. 3

3 Method

FM-ReID combines a holistic DINOv3 representation with selectively mined fine-grained features. Competing queries allocate patch tokens, and holistic and fine-grained representations receive ReID supervision for retrieval.

3.1 Overview of FM-ReID

  • The holistic branch combines the DINOv3 CLS token with average-pooled patch context, while CFM produces fine-grained representations from spatial patch tokens.Register tokens are discarded because they do not correspond to image locations.Source: §3.1 · p. 3

3.2 Competitive Fine-Grained Mining

  • Mining queries and a residual query compete for patch tokens, with query-wise normalization allocating each token among them.Position embeddings guide assignment through the keys, while the visual content being aggregated is unchanged.Source: §3.2 · p. 4
  • Above-prior selection suppresses assignments that do not exceed the uniform competition prior, then renormalizes each query over retained tokens for aggregation.Query supports may overlap and vary in size; tokens without above-prior mining assignments contribute to no fine-grained aggregate, and the residual output is excluded from retrieval.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.3 Optimization and Inference

  • Training supervises the holistic feature, holistic-context pairs for mined features, and a joint fusion that weights fine-grained blocks to prevent them overwhelming the holistic feature.The ReID objective combines label-smoothed identity classification and batch-hard triplet loss.Source: §3.3 · p. 5
  • At inference, individually L2-normalized holistic and fine-grained features are concatenated for retrieval, while competition maps and the residual output are discarded.With K = 5 and D = 768, the stored descriptor has 4,608 dimensions, compared with 768 for the holistic baseline.Source: Figure 2 · p. 4, §3.3 · p. 5

4 Experiments

Experiments evaluate FM-ReID on animal, person, and vehicle re-identification benchmarks. The setup includes a fish benchmark and uses benchmark-specific protocols and metrics.

4.1 Experimental Setup

  • Evaluation covers animal, person, and vehicle ReID, including WildlifeReID-10k, FM-FISH, four person benchmarks, and VeRi-776 and VehicleID for vehicles.Occluded-Duke evaluates retrieval under partial occlusion; WildlifeReID-10k covers 37 source datasets.Source: §4.1 · p. 6
  • FM-FISH comprises 45,632 images and 724 identity labels for matching visually similar fish, with source-specific splits.The self-collected subset uses unseen test identities, Wildlife subsets include known-identity queries, and MFT25 uses within-sequence temporal splits with track-based labels.Source: §4.1 · p. 6
  • The experiments fine-tune DINOv3 ViT-Base/16 with five mining queries and one residual query, using batch size 64 and final checkpoints without re-ranking.Images are resized to 256 × 256 for animal and vehicle ReID and 256 × 128 for person ReID; metrics follow the respective benchmark protocols, including mAP and CMC accuracy and WildlifeReID-10k's mTop-1, mTop-5, and BAKS.Source: §4.1 · p. 6

4.2 Evaluation on WildlifeReID-10k

FM-ReID leads the listed WildlifeReID-10k methods on mTop-1, mTop-5, and BAKS. Against Swin-Base, the strongest official baseline, it improves mTop-1 and BAKS by more than three percentage points.

4.2 Evaluation on WildlifeReID-10k

  • 85.03% mTop-1 and 92.07% mTop-5 put FM-ReID ahead of every listed WildlifeReID-10k method on both metrics.It also leads on BAKS.Source: §4.2 · p. 6
  • 83.33% BAKS makes FM-ReID the leader among the listed methods on that metric.Source: §4.2 · p. 6
  • 3.53 points on mTop-1 and 4.23 points on BAKS are FM-ReID’s gains over Swin-Base, the strongest official baseline.Source: §4.2 · p. 6

4.3 Evaluation on Person ReID Benchmarks

FM-ReID achieves the highest listed mAP on MSMT17, DukeMTMC-reID, and Occluded-Duke, with competitive results on Market-1501. On Occluded-Duke, it improves over CLIP-ReID in both reported metrics.

4.3 Evaluation on Person ReID Benchmarks

  • 92.3% mAP and 96.1% Rank-1 on Market-1501 are reported as competitive results; DukeMTMC-reID reaches 85.0% mAP and 92.2% Rank-1.Source: §4.3 · p. 6
  • 79.8% mAP and 91.3% Rank-1 on MSMT17 exceed EvoPrompt-ReID by 2.7 and 1.0 percentage points, respectively.FM-ReID has the highest listed mAP on MSMT17, DukeMTMC-reID, and Occluded-Duke.Source: §4.3 · p. 6, Table 2 · p. 7
  • 6.0 and 5.5 percentage points are FM-ReID’s gains over CLIP-ReID on Occluded-Duke mAP and Rank-1, respectively.FM-ReID achieves 65.5% mAP and 72.6% Rank-1 on this benchmark.Source: §4.3 · p. 7

4.4 Evaluation on Vehicle ReID Benchmarks

FM-ReID leads the listed methods on both vehicle benchmarks. It improves over EvoPrompt-ReID on VeRi-776 mAP and over Fast-ReID on VehicleID Rank-1 and Rank-5.

4.4 Evaluation on Vehicle ReID Benchmarks

  • 3.6 percentage points higher mAP than EvoPrompt-ReID accompanies FM-ReID’s 87.8% mAP and 98.3% Rank-1 on VeRi-776.FM-ReID leads the listed methods on this benchmark.Source: §4.4 · p. 7
  • 0.8 points on Rank-1 and 0.7 points on Rank-5 are FM-ReID’s VehicleID gains over Fast-ReID.FM-ReID reaches 87.4% Rank-1 and 98.6% Rank-5, leading the listed methods on VehicleID.Source: §4.4 · p. 7
  • FM-ReID leads the listed methods on both VeRi-776 and VehicleID.Source: §4.4 · p. 7

4.5 Evaluation on FM-FISH

On FM-FISH, FM-ReID outperforms TransReID and CLIP-ReID in the reported metrics. Ablations show gains from holistic aggregation, fine-grained mining, and retaining the residual query, while query-count sweeps support K = 5 as a compact choice.

4.5 Evaluation on FM-FISH

  • 11.9 and 9.8 percentage points over TransReID, and 13.7 and 8.3 points over CLIP-ReID, are FM-ReID’s FM-FISH gains on mAP and Rank-1, respectively.FM-ReID achieves 55.1% mAP and 63.5% Rank-1.Source: §4.5 · p. 7

4.6 Ablation Studies

  • Adding GAP to CLS improves WildlifeReID-10k mTop-1 by 0.99 points and MSMT17 mAP by 1.3 points; full CFM adds 0.75 and 2.3 points over CLS+GAP.Retaining the residual query adds 0.49 points in WildlifeReID-10k mTop-1 and 0.1 points in MSMT17 mAP.Source: §4.6 · p. 7, Table 4 · p. 8
  • 85.03% mTop-1 and 83.33% BAKS are reached at K = 5 on WildlifeReID-10k; MSMT17 mAP rises from 78.2% at K = 1 to 79.8% at K = 5.Source: §4.6 · p. 7, Figure 4 · p. 8
  • At K = 5, MSMT17 is within 0.2 points of the best mAP and 0.5 points of the best Rank-1 result.The paper describes this as a compact choice with competitive performance.Source: §4.6 · p. 8

4.7 Efficiency

FM-ReID adds modest inference cost relative to Baseline + GAP, while the CLS-to-GAP change adds parameters and peak memory without changing reported FLOPs.

4.7 Efficiency

  • 1.41% more FLOPs, 9.66% more parameters, and 9.28% more peak memory are reported for FM-ReID versus Baseline + GAP.The comparison is at the reported inference precision.Source: §4.7 · p. 8
  • Adding GAP to the CLS baseline adds 1.18M parameters and 4.5 MiB of peak memory, with unchanged FLOPs at the reported precision.Source: §4.7 · p. 8
  • TransReID uses patch stride 12 rather than 16, and its longer token sequence also contributes to its higher FLOPs.Source: §4.7 · p. 8

4.8 Qualitative Visualization

The qualitative examples show miner-specific spatial weighting patterns alongside the holistic representation, with responses ranging from localized to distributed and contextual.

4.8 Qualitative Visualization

  • Miner 3 concentrates on the ear region, while Miners 2 and 5 show more distributed responses over the body and surroundings.Source: §4.8 · p. 8
  • Miner 1 places greater emphasis on surrounding structures, illustrating contextual aggregation alongside the holistic representation.Source: §4.8 · p. 8
  • The additional miner features are combined with the holistic feature to enhance inter-identity separability in the fused feature space and support more accurate re-identification.Source: §4.8 · p. 8

5 Conclusion

The conclusion presents FM-ReID as a selective DINOv3-token approach that augments holistic representations. Results span supervised animal, person, and vehicle benchmarks, while transfer to unseen categories remains unevaluated.

5 Conclusion

  • FM-ReID extracts additional identity evidence from dense DINOv3 tokens through selective competition, using mining queries, a residual query, and above-prior selection to allow unclaimed evidence and unequal support sizes.The resulting features augment a holistic representation.Source: §5 · p. 8
  • FM-ReID exceeds the closed-set baselines reported by the WildlifeReID-10k benchmark paper, and within-backbone ablations evaluate its contribution beyond holistic pooling.Source: §5 · p. 8
  • Table 5 reports inference cost at 256×128, FP32, and batch size 1 on an RTX 5080.Source: Table 5 · p. 9
  • Figure 5 presents input images, holistic responses, and five miner attention maps for two panda identities; colors are normalized independently within each map.Source: Figure 5 · p. 9
  • Separate supervised experiments on animal, person, and vehicle benchmarks support cross-category applicability, but transfer to unseen object categories remains unevaluated.Source: §5 · pp. 8–9
Loading 2609.36560v1…