[Paper] Bi-C2R: Bidirectional Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification

Source: arXiv - 2512.25000v1

Overview

The paper introduces Bi‑C2R, a novel framework that enables lifelong person re‑identification (Re‑ID) without the costly “re‑indexing” step that traditionally forces all historic gallery images to be re‑processed after each model update. By making the feature representations of old and new models compatible with each other, Bi‑C2R lets systems continuously learn from new data while still being able to query against a static gallery—an important advance for privacy‑sensitive, large‑scale deployments.

Key Contributions

• New task definition – RFL‑ReID: Formalizes Re‑index‑Free Lifelong Re‑ID, where historic gallery features cannot be recomputed after model updates.
• Bidirectional Continuous Compatible Representation (Bi‑C2R): A dual‑direction compatibility learning scheme that aligns features from the current model to the old model and vice‑versa, ensuring seamless cross‑model retrieval.
• Theoretical compatibility guarantees: Provides analysis showing that the learned representations preserve similarity relationships across updates, mitigating catastrophic forgetting.
• Extensive empirical validation: Demonstrates state‑of‑the‑art performance on multiple benchmark datasets (e.g., Market‑1501, MSMT17) for both the new RFL‑ReID setting and the classic L‑ReID scenario.
• Efficiency gains: Eliminates the need to re‑extract features for millions of gallery images, cutting computational cost and storage overhead dramatically.

Methodology

1. Dual compatibility heads – The network attaches two projection heads to the backbone: one maps new model features into the old feature space, the other maps old features into the new space.
2. Bidirectional loss – During each incremental training phase, the model optimizes a combination of:
   • Cross‑model contrastive loss (aligns paired samples across time).
   • Standard Re‑ID classification & triplet losses (preserve discriminative power for the current data).
3. Memory‑free gallery update – After training, only the compatibility heads are applied to the existing gallery embeddings, producing updated vectors that are instantly usable with the latest query encoder. No raw images are stored or re‑processed.
4. Continual learning backbone – A lightweight rehearsal buffer (or regularization term) is used to keep old class prototypes, further reducing forgetting while keeping memory usage modest.

Results & Findings

• Near‑parity with full re‑indexing: Bi‑C2R closes >90% of the performance gap between the ideal (full re‑index) and the naïve “no‑update” baseline.
• Stable performance over many increments: After 10 sequential updates, the drop in mAP is <2%, indicating effective mitigation of catastrophic forgetting.
• Speed & storage: Updating the gallery via compatibility heads takes <0.5 s for a million embeddings, compared to >30 min for full feature extraction on the same hardware.

Practical Implications

• Privacy‑first deployments – Organizations can keep raw surveillance footage offline (or delete it) while still improving their Re‑ID models, because only compact embeddings need to be retained and updated.
• Cost‑effective scaling – Large‑scale retail or smart‑city systems can add new camera feeds or seasonal data without re‑processing the entire historic gallery, saving GPU hours and energy.
• Continuous improvement pipelines – Developers can integrate Bi‑C2R into CI/CD workflows: each model checkpoint automatically refreshes the gallery embeddings, enabling A/B testing of new architectures without downtime.
• Cross‑domain adaptability – Since compatibility is learned bidirectionally, the same framework can be reused when switching backbones (e.g., from ResNet‑50 to a lightweight MobileNet for edge devices) without breaking existing deployments.

Limitations & Future Work

• Compatibility head overhead – Adding two projection heads modestly increases model size and inference latency; future work could explore more lightweight alignment mechanisms.
• Rehearsal buffer dependence – The current approach still relies on a small buffer of past samples; eliminating this entirely would further reduce memory footprints.
• Domain shift robustness – While benchmarks show strong results, extreme domain shifts (e.g., night‑time vs. daytime cameras) may still degrade compatibility, suggesting a need for adaptive domain‑aware alignment.
• Extending beyond person Re‑ID – The authors plan to test Bi‑C2R on other retrieval tasks such as vehicle Re‑ID and product search, where similar re‑index‑free constraints exist.

Authors

• Zhenyu Cui
• Jiahuan Zhou
• Yuxin Peng

Paper Information

• arXiv ID: 2512.25000v1
• Categories: cs.CV
• Published: December 31, 2025
• PDF: Download PDF