Frustratingly Easy Person Re-Identification: Generalizing Person Re-ID in Practice

Problem setting

• DG Re-ID trains on labeled source data alone and tests on unseen datasets without adaptation, addressing the need to collect or annotate deployment-network data.Supervised Re-ID uses deployment-network annotations, while UDA still requires target data collection before adaptive training.Source: §1 · pp. 1–2

Approach

• DualNorm applies Instance Normalization in shallow layers to reduce style variation and feature Batch Normalization to address content-statistic differences.The paper associates style with color, lighting, clothing style, and related differences, and content with distance, viewpoint, and pose.Source: §1 · p. 2

Practical motivation

• The authors emphasize simplicity, stability, and efficiency as practical advantages over elaborate alternatives.They argue ease of implementation and efficiency matter for practical use.Source: §1 · p. 2

Contributions

• DualNorm is presented as a strong DG Re-ID baseline that needs no target-domain information during training and works as an out-of-the-box feature extractor for novel camera networks.The paper also claims it surpasses most target-domain-supervised methods and all methods that do not learn on target data.Source: §1 · p. 2

Domain adaptation and generalization

• UDA methods can reduce annotation costs but require prior target-domain data collection, whereas DG seeks models robust to shifts between training and testing domains.In Re-ID, UDA approaches include image synthesis and source-target alignment.Source: §2 · p. 3

Related DG and normalization methods

• The paper contrasts its minor modification of standard Re-ID architectures with DIMN’s meta-learning and dynamic model synthesis, which it describes as complicated, slow, and less amenable to runtime scalability modifications.The authors characterize their contribution as integrating IN in MobileNetV2 bottlenecks and BN-based feature normalization into an end-to-end DG network.Source: §2 · p. 3

3 Method

• The setup assumes labeled source datasets with disjoint identity label spaces, whose union forms the combined identity space.The method specifies K source datasets and N total identities across them.Source: §3 · pp. 3–4

A Strong Cross-domain Baseline

• The baseline aggregates labeled images from multiple source domains to train one identity classifier, then uses the learned model as an off-the-shelf feature extractor on unseen identities.It uses MobileNetV2 and retrieves target-set matches from 1280-dimensional pooling-layer activations using Euclidean distance.Source: A Strong Cross-domain Baseline · p. 4

3.1 Style Normalization

• Instance Normalization is added to shallow MobileNetV2 bottlenecks to filter instance-specific style variation; the method avoids applying it in deep layers because that may remove discriminative information and degrade classification.The paper describes image style as encoded in first-order feature-map statistics.Source: §3.1 · pp. 4–5

3.2 Content Normalization

• After global average pooling, Batch Normalization of the penultimate features addresses content-statistic differences; its bias is fixed to zero and the following classifier bias is removed.The paper links content differences to camera angle, body size, viewpoint, and pose, and says the added normalization layers introduce negligible parameters and computation cost.Source: §3.2 · pp. 5–6

4.1 Datasets and Settings

• The multi-source training set combines five datasets, totaling 18,530 training identities and 121,765 images, for evaluation on four target datasets.The sources are Market-1501, DukeMTMC-reID, CUHK02, CUHK03, and PersonSearch; testing follows standard single-shot protocols and averages results over 10 random splits.Source: §4.1 · p. 6

4.2 Comparisons Against State-of-the-Art

• DualNorm consistently outperforms most unsupervised and supervised competitors, including the purpose-designed DG method DIMN, without using target images or labels.Table 1 reports Rank 1 accuracy on VIPeR, PRID, GRID, and i-LIDS; supervised and unsupervised methods use target data, unlike DG methods.Source: Table 1 · p. 7, §4.2 · pp. 7–8

4.3 Additional Analysis

• Adding IN and FN both improves AGG, and their combination yields further gains; IN in early layers appears more beneficial than FN alone.The reported Rank 1 rate increases are 11.8%, 33.2%, 12.8%, and 8.5%, respectively; the added normalization layers introduce few parameters.Source: §4.3 · p. 8, Table 2 · p. 8
• With FN removed, adding more IN in early MobileNetV2 layers improves performance, while deep-layer IN is detrimental and Conv1 IN is critical.Table 3 analyzes IN placement by rank 1 accuracy; the comparison of placements ‘1-6’ and ‘2-6’ supports the Conv1 finding.Source: §4.3 · p. 8, Table 3 · p. 9
• DualNorm performs consistently better than the compared feature-normalization alternatives, while methods such as HAFN tend to reduce overall performance.The authors suggest additional hyperparameters and l2-normalized feature embeddings may account for the drop, and may not cooperate well with early IN.Source: §4.3 · pp. 8–9, Table 4 · p. 9
• On ResNet50, DualNorm improves over the vanilla backbone and outperforms IBN-a and IBN-b, but the effect of Conv1 IN differs from MobileNetV2.Early-layer IN helps and deep-layer IN hurts on ResNet50; unlike the MobileNetV2 result, Conv1 IN is detrimental in ResNet50.Source: §4.3 · p. 9, Table 5 · p. 10
• Within-dataset Rank 1 accuracy increases by 5.4% on Market-1501 and 6.2% on DukeMTMC-reID.The authors report smaller gains than in cross-dataset experiments, where train-test domain bias is more pronounced.Source: Table 6 · p. 10, §4.3 · p. 10

• The proposed normalization-based baseline surpasses contemporary supervised, unsupervised, and purpose-designed DG methods for multi-source cross-domain Re-ID.Source: §5 · p. 10
• Combining Instance and Feature Normalization alleviates style and content biases between domains, promoting generalization and transferability of deep Re-ID models.Source: §5 · p. 10
• The paper offers a fast, simple, effective baseline for engineers implementing Re-ID and for future research.Source: §5 · p. 10