BagDINO: Multi-View Baggage Re-Identification with DINOv3

• Tag-based tracking does not directly support image-based recognition when tags are missing, corrupted, or unavailable during inspection.Existing tracking records checkpoints including Acceptance, Load, Transfer, and Arrival.Source: §I · p. 1
• A small set of bag images captured during controlled registration can form a gallery for later matching of airport-camera query images.The registration images use stable viewpoint, lighting, and background conditions.Source: §I · p. 1
• Baggage ReID must distinguish near-identical items using fine-grained cues that can be obscured by viewpoint changes, occlusion, blur, illumination, and clutter.Examples of instance-specific cues include stickers, surface wear, and small damages.Source: §I · p. 1
• The proposed multi-camera instance-retrieval pipeline combines DINOv3 ViT-L/16 representations with a Torchreid-style BNNeck head and LoRA updates to selected projections while freezing backbone weights.The adaptation configuration evaluates LoRA updates to qkv, proj, fc1, and fc2.Source: §I · pp. 1–2
• LoRA reached 0.88 mAP and 0.87 Rank-1 on MVB, compared with 0.57 mAP and 0.60 Rank-1 for the frozen setting.The study also compared partial and full fine-tuning under a shared training recipe.Source: §I · p. 2

• MVB contains 4,519 identities and 22,660 images from real multi-camera airport environments, with domain shift and substantial viewpoint, illumination, occlusion, blur, and clutter variation.Baggage identities often have highly similar appearances, while each identity can vary substantially across views.Source: §II · p. 2
• Classical MVB approaches adapt CNN-based ReID recipes, while attention, multi-granularity features, and multi-view losses add local cues and cross-view discrimination.Mazzeo et al. combine global and local features; QuadNet uses a multi-view quadruplet loss and view-aware local features.Source: §II · p. 2
• Cross-domain methods combine domain-robust feature learning, hard-sample mining, and camera-aware re-ranking to improve MVB mAP and Rank-1.CDR-CARNet is cited as an example of this direction.Source: §II · p. 2
• RGViT aggregates local evidence into regional tokens and models spatial relations, reporting 88.7% Rank-1 and 89.6% mAP on MVB.Its quadruplet objective is designed for near-duplicate instances.Source: §II · p. 2
• The paper argues that person-ReID adaptations may miss baggage-specific cues because baggage has weaker structural priors, similar categories, and frequent handling-related occlusions or deformations.It identifies specialized feature learning for global structure and subtle local differences as a need.Source: §II · p. 2

• BagDINO aims to improve robustness to viewpoint variation, background clutter, and visual similarity across baggage identities.The stated goal is more reliable instance matching in real-world airport scenarios.Source: §III · p. 2
• The backbone uses the DINOv3 foundation model with parameter-efficient fine-tuning through Low-Rank Adaptation.Source: §III · p. 2
• The ReID head produces final embeddings from the backbone patch token, and training combines triplet loss with cross-entropy loss.Source: §III · p. 2

A. Backbone and LoRA Fine-Tuning

• DINOv3 supplies the visual backbone, using self-supervised pretraining and representations described as transferable across visual tasks.The pipeline uses its [CLS] token for a global image representation.Source: A. Backbone and LoRA Fine-Tuning · p. 2, Figure 1 · p. 3
• LoRA is chosen because the MVB training split is relatively small and full fine-tuning a large ViT is computationally costly.Source: A. Backbone and LoRA Fine-Tuning · p. 2
• LoRA adds trainable low-rank matrices to selected Transformer projections while freezing pretrained weights; its zero-initialized update uses B = 0 and random Gaussian A [6].The rank r controls trainable parameter count, with r ≪ min(d, k), and α governs adaptation magnitude.Source: A. Backbone and LoRA Fine-Tuning · p. 3
• Adapters target qkv, proj, fc1, and fc2, while a BNNeck head is trained with identity cross-entropy and batch-hard triplet loss.At inference, the classifier is removed and normalized embeddings rank gallery items, for example by cosine similarity.Source: A. Backbone and LoRA Fine-Tuning · p. 3, Figure 1 · p. 3

B. Re-Identification Head

• The head uses the [CLS] token as a global image descriptor before passing it through the re-identification head.The token aggregates information from spatial positions through self-attention.Source: B. Re-Identification Head · p. 3
• At inference, the classifier is discarded and gallery images are ranked by cosine similarity to the query embedding.The backbone and BNNeck embed both query and gallery images to produce the match list.Source: B. Re-Identification Head · p. 3, Figure 2 · p. 3
• The BNNeck separates features used for metric learning from those used for identity classification.A linear classifier with one output per training identity is used only during training.Source: B. Re-Identification Head · p. 3

C. Loss Function

• Identity loss applies cross-entropy with label smoothing to post-BN features.Label smoothing prevents overconfident predictions on training identities.Source: C. Loss Function · p. 4
• Batch-hard triplet loss selects the hardest same-identity positive and different-identity negative for each mini-batch anchor.The loss uses Euclidean distance and a margin.Source: C. Loss Function · p. 4
• The total objective is a weighted sum of identity classification loss and metric-learning loss.λ controls the metric-learning term’s relative contribution.Source: C. Loss Function · pp. 3–4

A. MVB Dataset

• MVB contains 4,519 identities and 22,660 images spanning diverse baggage types and two acquisition points with domain gaps and intraclass variation.Variations include viewpoint, lighting, occlusions, motion blur, and background clutter.Source: A. MVB Dataset · p. 4

B. Experimental Setup

• Evaluation follows the official MVB train/query/gallery splits and reports mAP and CMC at Rank-1, Rank-5, and Rank-10.The shared recipe uses identity classification and batch-hard triplet loss, with images resized to 256 × 256 and training for 80 epochs.Source: A. MVB Dataset · p. 4, B. Experimental Setup · p. 4

A. MVB Dataset

• Example query-gallery pairs illustrate baggage heterogeneity and variability across acquisition conditions.The benchmark includes BHS/conveyor-belt and checkpoint views.Source: Figure 3 · p. 4, A. MVB Dataset · p. 4

B. Experimental Setup

• LoRA updates are injected into the backbone’s qkv, proj, fc1, and fc2 projections.The low-rank updates use rank 16 and α = 32.Source: B. Experimental Setup · p. 4

C. Progressive Study

• The frozen DINOv3 backbone baseline achieves 0.57 mAP and Rank-1/5/10 of 0.60/0.80/0.87, outperforming the original MVB baseline.Only the ReID head is trained in this setting.Source: C. Progressive Study · p. 4
• LoRA reaches 0.88 mAP and 0.87 Rank-1, ahead of frozen (0.57 mAP, 0.60 Rank-1), partial fine-tuning (0.83 mAP, 0.83 Rank-1), and full fine-tuning (0.85 mAP, 0.86 Rank-1).LoRA also achieves Rank-5/10 of 0.96/0.98; partial fine-tuning reaches 0.95/0.97 and full fine-tuning 0.95/0.96.Source: C. Progressive Study · pp. 4–5
• Partial fine-tuning achieves 0.83 mAP and Rank-1/5/10 of 0.83/0.95/0.97, but the paper suggests more data may be needed to adapt without degrading generalization.In this configuration, 75% of the backbone remains frozen and the rest is updated with a conservative learning rate and stronger weight decay.Source: C. Progressive Study · p. 5
• Full fine-tuning achieves 0.85 mAP and Rank-1/5/10 of 0.86/0.95/0.96; the paper concludes LoRA is more stable and effective given limited training data and high visual similarity.All backbone parameters are unfrozen under a conservative optimization regime.Source: C. Progressive Study · p. 5

D. Error Analysis

• High-confidence retrieval errors occur when different baggage identities have near-duplicate shape, material, and color cues, producing close embedding-space neighbors.The reported examples show the query, its true gallery match, and a top-ranked incorrect retrieval with similarity ≳78%.Source: D. Error Analysis · p. 5, Figure 4 · p. 5

E. Results Summary

• LoRA-based adaptation yields 0.88 mAP and 0.87 Rank-1, the best overall trade-off among the tested MVB settings.Frozen, partial, and full fine-tuning remain competitive but do not match LoRA under the same training recipe.Source: E. Results Summary · p. 5, Table I · p. 6
• BagDINO + LoRA is competitive with prior MVB work and achieves results close to the strongest published methods while using parameter-efficient adaptation.Table II compares mAP and Rank-1 across representative methods.Source: E. Results Summary · p. 5, Table II · p. 6
• The method comparison reflects different design trade-offs: BagDINO uses limited low-rank updates with the DINOv3 backbone frozen, while some methods add architectural modules such as RGViT.The paper cautions that benchmark-score differences are not the only distinction between approaches.Source: E. Results Summary · p. 5
• Across the baggage ReID methods considered on Market-1501, BagDINO + LoRA achieves the highest mAP and Rank-1.The evaluation is presented as out-of-domain validation, not an in-domain comparison with MVB.Source: E. Results Summary · p. 6, Table III · p. 6

V. CONCLUSIONS

• 0.88 mAP and 0.87 Rank-1: LoRA adaptation provided the strongest overall results among tested configurations on MVB.The training regime remained stable under limited training data.Source: §V · p. 6
• Limited gallery images per identity in the test split make purity at fixed top-k inherently conservative.Source: §V · p. 6
• Remaining failures typically involve visually near-duplicate baggage; future evaluations could use larger registration templates and added disambiguation mechanisms.The proposed mechanisms include stronger fine-grained cues, re-ranking, or auxiliary supervision without compromising retrieval efficiency.Source: §V · p. 6