Lifelong Person Re-Identification via Adaptive Knowledge Accumulation

1. Introduction

• ReID methods trained on a fixed dataset are limited in settings where data continually arrives from different domains.The introduction gives multi-crossroads surveillance as an example of a setting requiring incremental or lifelong learning.Source: §1 · p. 1
• LReID requires incrementally accumulating knowledge across seen domains and adapting it for tests on both seen and unseen domains.The task also considers generalization to unseen classes and fine-grained appearance differences.Source: §1 · p. 1
• AKA represents transferable knowledge as a learnable knowledge graph and uses graph operations to transfer and update knowledge for current samples.Plasticity and stability losses balance improving learned representations with addressing forgetting.Source: §1 · pp. 1–2
• The paper provides a new benchmark and evaluation protocols for LReID.The benchmark includes seen and unseen subsets.Source: §1 · p. 2

2.1. Person Re-identification Setups

• Existing ReID setups include fully supervised, unsupervised domain adaptation, pure unsupervised, and domain generalization, but do not address lifelong learning.Table 1 compares these setups with LReID; the supplied text does not give its cell-level comparisons.Source: §2.1 · p. 2, Table 1 · p. 2
• Online person ReID and continual representation learning focus on intra-domain rather than interdomain incremental learning.The cited works are and [48], respectively; their training-to-test domain gaps are described as relatively narrow.Source: §2.1 · p. 2

2.2. Lifelong Learning

• LReID combines catastrophic forgetting with ReID-specific challenges: many identities, imbalanced few-shot classes, fine-grained distinctions, and unseen test classes.The passage notes that remembering seen classes is beneficial for generalizing to newly unseen classes.Source: §2.2 · p. 2

3.1. Problem Definition and Formulation

• LReID evaluates retrieval on encountered domains and generalization on new unseen domains, while training and testing classes are disjoint.Data from previous domains are unavailable after the corresponding training step.Source: §3.1 · p. 3

3.2. Baseline Approach

• The baseline maps images to class-confidence scores and trains its feature extractor and classifier with cross-entropy loss.The network is defined as a classifier applied to the feature extractor.Source: §3.2 · p. 3
• The baseline adds knowledge distillation to mitigate forgetting on previous domains, with a trade-off factor balancing distillation and cross-entropy losses.The distillation teacher parameters are copied before current-step training.Source: §3.2 · p. 3

4.1. Knowledge Representation

• AKA represents current-batch relationships with an instance similarity graph and accumulated knowledge with a learnable knowledge graph.The two graphs use independent learnable weights because they organize and transform knowledge differently.Source: §4.1 · pp. 3–4

4.2. Knowledge Operation

• AKA links the current-sample and knowledge graphs, propagates relevant knowledge with a graph convolutional network, and sums acquired representations with input features.The cross-graph weights are based on Euclidean distances, and the propagated features come from the current-sample rows of the joint graph.Source: §4.2 · p. 4
• Plasticity guides acquired representations toward better generalization, while stability loss limits large knowledge-graph vertex movements to reduce overfitting and forgetting.The stability term constrains vertices toward their initial parameters as the graph updates.Source: §4.2 · pp. 4–5

4.3. Optimization

• Optimization combines base, plasticity, and stability losses, with trade-off factors balancing new-knowledge learning against preserving accumulated knowledge.Training uses mini-batch SGD and detaches gradients between the feature and knowledge parameters.Source: §4.3 · p. 5

4.4. Discussion

• AKA uses non-parameterized knowledge transformation and parameterized knowledge representation, while gradient detachment supports domain-specific knowledge transfer.The discussion also describes preserving vertex topology while flexibly adapting transformation weights to represent new knowledge.Source: §4.4 · p. 5

5.1. Implementation Details

• Experiments use a ResNet-50 feature extractor and one GCN layer, with five repeated runs reporting means and standard deviations.The training batches contain 32 identities with four images each; the learning rate is reduced at epochs 25 and 35.Source: §5.1 · p. 5

5.2. New Benchmark for LReID

• The benchmark includes seen and unseen subsets, about 14K classes, and testing on novel identities even in unseen domains.The seen subset uses five domain-incremental training steps across five datasets; the unseen subset merges seven ReID datasets for evaluation.Source: §5.2 · pp. 5–6, Table 2 · p. 6

5.3. Seen-domain Non-forgetting Evaluation

• For Order-1, AKA’s seen-domain accuracy degradation is 17.5%/14.7% in mAP/Rank-1; for Order-2, it is 29.3%/27.9%.For Order-1, AKA outperforms SFT by around 30% in Rank-1 and CRL by 6% in mAP; training order significantly affects forgetting.Source: Table 3 · p. 7, §5.3 · p. 7, Figure 3 · p. 7

5.5. Ablation Study

• AKA improves mAP and less-forgetting score by 6% over LwF; plasticity loss benefits unseen domains, while plasticity and stability losses are complementary.The hyperparameter analysis reports best balance between seen and unseen domains at λp = 1 and λs = 5 × 10−4; the knowledge-graph vertex count is not sensitive, and unseen-domain mAP increases as it grows.Source: §5.5 · pp. 7–8, Table 5 · p. 8

5.4. Unseen-domain Generalising Evaluation

• AKA ranks best among the compared methods, with 31.8% averaged mAP on seen domains and 44.3% averaged mAP on unseen domains.Methods without knowledge distillation perform better at step 2 but do not continue accumulating knowledge for improved generalization; the pattern also appears under Order-2.Source: §5.4 · p. 7, Table 4 · p. 8, Figure 4 · p. 8

• LReID requires models to leverage previous knowledge to improve generalization on both seen and unseen domains.Source: §6 · p. 8
• Figure 5 examines gradient detachment using VS and ¯VS cosine similarity, and studies λp, λs, and NK.Source: Figure 5 · p. 8
• AKA outperforms competitors in both mitigating forgetting on seen domains and generalizing on unseen domains.Source: §6 · p. 8
• AKA preserves knowledge learned from previous domains while adaptively propagating it to improve learning on new domains.Source: §6 · p. 8