How Bi‑C2R Achieves Re‑indexing‑Free Lifelong Person Re‑identification

Introduction

Lifelong person re‑identification (L‑ReID) aims to continuously learn from streaming pedestrian data without degrading performance on previously seen identities. Traditional L‑ReID requires costly feature reconstruction of the entire gallery and raises privacy concerns because original images cannot be stored. The authors propose a re‑indexing‑free solution that updates gallery features directly.

Proposed Method: Bi‑C2R

Bi‑C2R (Bidirectional Continual Compatible Representation) consists of four key components:

Bidirectional Continual Transfer Network (BiCT‑Net)

Bidirectional Compatible Distillation (BiCD)

Bidirectional Anti‑forgetting Distillation (BiAD)

Dynamic Feature Fusion (DFF)

Bidirectional Continual Transfer Network

BiCT‑Net is a dual‑stream network that contains a Knowledge Capture module and a Forward Mapping module. The Knowledge Capture module computes a capture probability for each new‑knowledge prototype using a three‑layer fully‑connected network followed by a Softmax and L2‑normalisation, then aggregates the prototypes by weighted sum. The Forward Mapping module reduces the channel dimension with a fully‑connected layer, applies Batch Normalisation and PReLU, and restores the original dimension. A learnable adaptive weight balances the outputs of the two branches, and a residual connection mitigates gradient vanishing. The transferred feature 𝑓′ is obtained by:

f′ = α·f_capture + (1‑α)·f_forward + f_residual

Bidirectional Compatible Distillation

BiCD aligns the transferred feature space with the new model’s feature space using an L2 loss. It then constructs similarity matrices for the source and transferred features, removes intra‑identity similarities, and normalises inter‑identity similarities. A Kullback‑Leibler (KL) divergence loss preserves the relational structure: ℒ_KL = Σ_i KL( Ŝ_i^src ‖ Ŝ_i^trans ) where 𝑆̂ denotes the normalised similarity matrix for sample i.

Bidirectional Anti‑forgetting Distillation

BiAD keeps the discriminative power of old features. A classification head provides identity probability distributions 𝑝_old from old features. To counter domain shift, the old distribution is imposed on the transferred features, and an anti‑forgetting loss encourages the transferred features to retain the same class‑wise statistics:

ℒ_AF = ‖μ_new – μ_old‖² + ‖σ_new – σ_old‖²

A consistency term further balances the discriminative ability of new and transferred features.

Dynamic Feature Fusion

Because domain gaps vary, DFF adaptively fuses the current stage’s new features with the updated gallery features. The fusion weight is derived from the prediction difference between the old and new models: γ = sigmoid(‖p_new – p_old‖) The final gallery feature at stage t is:

G_t = γ·F_t^new + (1‑γ)·F_t^old

Experiments

The authors evaluate Bi‑C2R on five widely used person Re‑ID datasets (Market‑1501, DukeMTMC‑ReID, CUHK‑03, MSMT17, CUHK‑SYSU) under two training orders (Order‑1 and Order‑2). Both lifelong Re‑ID (L‑ReID) and re‑indexing‑free lifelong Re‑ID (RFL‑ReID) are tested using mAP and Rank‑1 metrics.

Results show that Bi‑C2R outperforms prior methods by 2.7%–3.3% absolute gain in mAP/Rank‑1 on RFL‑ReID and by 0.2%–0.3% on traditional L‑ReID. The improvements are attributed to (1) bidirectional transfer that preserves old knowledge while acquiring new discriminative features, and (2) dynamic fusion that mitigates feature drift caused by exponential moving average updates.

Conclusion

Bi‑C2R provides a comprehensive solution for re‑indexing‑free lifelong person re‑identification. By jointly addressing feature compatibility, knowledge retention, and domain adaptation through bidirectional transfer, compatible and anti‑forgetting distillation, and dynamic fusion, the method achieves superior accuracy in both privacy‑preserving and standard lifelong scenarios.