Core Concepts
Proposing a methodology for lifelong person re-identification that ensures backward-compatibility with old models, improving performance.
Abstract
The paper introduces a novel approach to lifelong person re-identification (LReID) that focuses on maintaining model compatibility with previously trained models. By incorporating backward-compatibility, the proposed method addresses the issue of catastrophic forgetting and reduces the need for time-consuming backfilling of features during inference. The methodology involves cross-model compatibility loss and knowledge consolidation based on part classification to ensure shared representation across datasets. Experimental results demonstrate significant improvements in performance compared to existing methods, especially in practical scenarios. The proposed evaluation methodology considers all gallery and query images simultaneously, providing a more comprehensive assessment of performance.
Stats
Training Images: 12,936 (Market1501), 16,522 (DukeMTMC), 15,088 (CUHK-SYSU), 32,621 (MSMT17)
ID: 751 (Market1501), 702 (DukeMTMC), 5,532 (CUHK-SYSU), 1,041 (MSMT17)
Query Images: 3,368 (Market1501), 2,228 (DukeMTMC), 2,900 (CUHK-SYSU), 11,659 (MSMT17)
Quotes
"We propose a more practical methodology for performance evaluation where all the gallery and query images are considered together."
"Experimental results demonstrate that the proposed method achieves a significantly higher performance of the backward-compatibility compared with the existing methods."