Adaptive Exploration for Unsupervised Person Re-Identification

• Directly applying a model to an unseen dataset often yields poor accuracy, while collecting large-scale labeled target data is costly and time-consuming.Source: §1 · p. 2
• Target-only re-ID trains on unlabeled target data, while domain-adaptive re-ID also uses labeled source data; AE is mainly discussed under the latter protocol.Source: §1 · p. 2
• A non-parametric classifier with feature memory separates all target images while pulling similar images together by treating them as one class.Source: §1 · p. 2
• AE begins with a small set of reliable similar samples selected by a similarity threshold, then expands selection as the model improves.Source: §1 · pp. 2–3
• A balance strategy counteracts neighborhood imbalance, in which images with many neighbors can attract more images, including unreliable ones.Source: §1 · p. 3
• AE significantly improves unsupervised re-ID accuracy on three large-scale datasets under both target-only and domain-adaptive protocols.Source: §1 · p. 3

2 RELATED WORK

• The review distinguishes supervised and unsupervised person re-ID, with unsupervised work divided into target-only and domain-adaptive approaches.Source: §2 · p. 3

2.1 Supervised Person Re-identification

• Supervised re-ID work spans siamese and classification models, losses and augmentation, plus methods addressing pose, viewpoint, background, and occlusion variation.Source: §2.1 · pp. 3–4

2.2 Target-only Person Re-identification

• Target-only methods use unlabeled target data; examples include feature memory with online instance matching and bottom-up clustering.Source: §2.2 · p. 4

2.3 Domain Adaptive Person Re-identification

• Domain-adaptive re-ID uses labeled source and unlabeled target data; unlike fixed-count neighborhood invariance, AE selects neighborhoods adaptively and balances training.Source: §2.3 · p. 4

3 ADAPTIVE EXPLORATION METHOD

• AE combines supervised source-image training with target-domain learning that maximizes distances between all target images and minimizes distances between similar ones.The framework uses a non-parametric classifier with feature memory for target images.Source: §3 · pp. 4–5, Figure 1 · p. 5

3.1 Parametric Classifier for Source Domain

• Source images are trained with supervised learning and a parametric classifier over labeled person identities.The classifier is optimized using source-domain labels.Source: §3.1 · p. 6

3.2 Non-parametric Classifier with Feature Memory

• The non-parametric classifier treats target images as individual classes or shared classes, using a memory that stores target features and updates during training.Treating each image as its own class supports separation, while grouping similar images supports attraction.Source: §3.2 · p. 6

3.3 Learning with Adaptive Selection

• Adaptive selection treats each target image as an individual class for separation, then uses threshold-selected neighborhoods as the same class to bring similar images closer.Pairs are neighborhoods when their distance is below λ.Source: §3.3 · pp. 7–8

3.4 Learning with Balance

• Unequal neighborhood counts can let large groups attract more images, so a balance term penalizes high-neighborhood-count images more heavily.The penalty is 1/(∥v_i∥_1 log(∥v_i∥_1)); images with only themselves are excluded from training.Source: §3.4 · pp. 8–9

3.5 Optimization Procedure

• Optimization alternates between selecting reliable neighborhoods and training the model, then updates feature memory; target-only training omits source loss.For target-only re-ID, ξ is set to 1.Source: §3.5 · p. 10

4 EXPERIMENT

• The evaluation reports accuracy on Market-1501, DukeMTMC-reID, and MSMT17.Market-1501 and DukeMTMC-reID are described as widely used by existing methods.Source: §4 · p. 10

4.1 Datasets and Settings

• Market-1501 contains 32,688 images across 1,501 identities captured by six cameras.Its training split has 12,936 images and 751 identities; gallery and query use the same 750 identities.Source: §4.1 · p. 10
• DukeMTMC-reID contains 36,411 images across 1,812 identities collected from eight cameras.Training uses 16,522 images and 702 identities; gallery and query use the same 702 identities.Source: §4.1 · p. 10
• MSMT17 contains 126,441 images across 4,101 identities captured by 15 cameras.Training uses 32,621 images and 1,041 identities; gallery and query use the same 3,060 identities.Source: §4.1 · p. 10
• Evaluation reports rank1, rank5, rank10, and mean average precision (mAP) with single-query testing.Source: §4.1 · p. 10

4.2 Implementation Details

• The model uses ImageNet-pretrained ResNet-50 features, with a 4,096-dimension training representation and 2,048-dimension Pool-5 testing features.A 4,096-dimension fully connected layer follows Pool-5 for training.Source: §4.2 · p. 10
• Source images use a parametric classifier, while target images use a feature-memory non-parametric classifier.The first two ResNet-50 layers are fixed during training.Source: §4.2 · p. 11
• Training uses random flip, crop, and erasing, resizes images to 256 × 128, and samples 128 images from each domain per iteration.CamStyle is also used to reduce target-domain camera variance.Source: §4.2 · p. 11
• SGD training uses a learning rate of 0.01 for base layers and 0.1 for other layers for 40 epochs, then divides both rates by 10 for 20 epochs.Similar-image minimization starts after five epochs; cosine similarity selects neighborhoods above threshold λ.Source: §4.2 · p. 11
• The similarity threshold λ and temperature τ are set to 0.55 and 0.05, respectively.The memory update rate µ rises linearly from 0 to 0.4 on Market-1501 and from 0 to 0.5 on DukeMTMC-reID; δ and ξ are 3.5 and 0.6.Source: §4.2 · p. 11

4.3.1 Exploration for Update Rate

• Varying µ outperforms constant µ, with peak accuracy at 0→0.4 for Market-1501 and 0→0.5 for DukeMTMC-reID.Accuracy first rises and then falls as the varying rate increases.Source: §4.3.1 · p. 11

4.3.2 Learning with Adaptive Exploration.

• λ=0.55 gives the highest accuracy on both Market-1501 and DukeMTMC-reID, and adaptive selection outperforms top-k across cases.At λ=0.55, the last-epoch average is 18.3 neighborhoods, close to Market-1501's 17.2 images per identity.Source: §4.3.2 · pp. 11–12

4.3.3 Learning with Balance.

• With CamStyle augmentation for DukeMTMC-reID→Market-1501, balance improves rank-1 accuracy by +15.0% and mAP by +22.9%.The authors report that learning with balance significantly improves accuracy.Source: §4.3.3 · pp. 12–13

4.3.4 Comparison between Transfer Learning and Target-only Learning.

• For DukeMTMC-reID→Market-1501, AE gains 35.2% rank-1 accuracy and 38.0% mAP over source-only; for Market-1501→DukeMTMC-reID, it gains 37.6% rank-1 accuracy and 20.8% mAP.These results compare models trained with both source and target data against the source-only baseline.Source: §4.3.4 · p. 13
• On Market-1501, target-only AE reaches 77.5% rank-1 accuracy and 54.0% mAP, 4.1% and 4.0% below transfer, respectively.Target-only training uses target data without source data.Source: §4.3.4 · p. 13

4.3.5 Hyper-parameter Analysis.

• Accuracy is higher at τ=0.05; ξ endpoints reduce it while ξ∈[0.3,0.9] is stable, and δ∈[3.0,4.0] performs better.Each hyperparameter was varied while the others were fixed.Source: §4.3.5 · pp. 13–14, Figure 4 · p. 14

4.3.6 Comparison between Triplet and Contrastive Loss.

• Triplet and contrastive losses fail to reach AE's competitive accuracy; 2048- and 4096-dimensional features perform similarly, but 4096 uses more flops and parameters.The 2048-dimensional features come from Pool-5; the 4096-dimensional features come from the fully connected layer.Source: §4.3.5 · p. 14, Table 2 · p. 14, Table 3 · p. 14

4.4 Comparison with State-of-the-art Methods

• AE outperformed state-of-the-art comparisons in both domain-adaptive and target-only re-ID on Market-1501 and DukeMTMC-reID, and achieved competitive target-only results on MSMT17.The comparisons include ten domain-adaptive methods, four target-only methods on Market-1501 and DukeMTMC-reID, and three methods on MSMT17.Source: §4.4 · pp. 15–16
• Against ECN, AE raises rank-1 accuracy by 6.5% on Market-1501 and 4.6% on DukeMTMC-reID.The comparisons are reported in Table 4.Source: §4.4 · p. 15
• Against BUC, target-only AE raises rank-1 accuracy by 11.3% on Market-1501 and 15.8% on DukeMTMC-reID.On Market-1501, AE also records 77.5% rank-1 and 54.0% mAP; on DukeMTMC-reID, 63.2% rank-1 and 39.0% mAP.Source: §4.4 · p. 15, Table 5 · p. 16
• On MSMT17, target-only AE records 26.6% rank-1 accuracy and 8.5% mAP, gains of 15.1% and 5.1% over BUC, respectively.Table 6 compares AE with two domain-adaptive methods and one target-only method.Source: §4.4 · pp. 15–16, Table 6 · p. 16

4.5.1 Effectiveness of Learning with Balance.

• With balance, the two DukeMTMC-reID images select similar neighborhood counts whose identities match theirs; without balance, one selects too many and both groups include incorrect identities.Figure 5 contrasts selections with and without the balance term.Source: §4.5.1 · p. 16, Figure 5 · p. 17

4.5.2 Effectiveness of the AE Method.

• In AE's feature map, some different-color points remain together, consistent with the model sometimes classifying similar people as one identity without labels.Figure 6(d) shows the projected feature distribution.Source: §4.5.2 · pp. 16–17, Figure 6 · p. 17
• Compared with source-only, BUC, and ECN, AE's feature map shows same-identity points closer together and fewer different-identity points grouped by mistake.Figure 6 visualizes 1,926 Market-1501 gallery images from 100 identities; each point is an image and each color an identity.Source: §4.5.2 · pp. 16–17, Figure 6 · p. 17

5 CONCLUSION

• AE uses a non-parametric classifier with feature memory to push all target-image features apart while bringing similar target images together.The memory stores target-image features and is updated after each iteration.Source: §5 · p. 18
• A similarity threshold selects reliable similar images, while a balance term discourages images with many neighborhoods from attracting others.The stated aim is for images to select a balanced and reasonable number of neighborhoods.Source: §5 · p. 18
• With adaptive selection and balancing, AE achieves competitive accuracy in both target-only and domain-adaptive re-ID.Source: §5 · p. 18