Bridging the Short-Term and Long-Term Gap: A Cross-Task Continuous Learning Person Re-Identification Problem

Person Re-IDentification (ReID) plays an important role in the application of intelligent security systems. Significant advancements have been made in Short-Term (ST) ReID tasks, where person appearances remain relatively constant. Similarly, progress has been achieved in Long-Term (LT) ReID tasks, which are characterized by drastic changes in person appearances. However, a vital yet overlooked challenge in real-world ReID is maintaining the continuity of the retrieval process. When transitioning from ST to LT ReID, simply discarding the ST model in favor of the LT model can lead to a severe performance drop during the early retrieval stage, due to the loss of valuable appearance-related knowledge in the ST model. Conversely, discarding the LT model and relying solely on the ST model may significantly reduce accuracy during the extended retrieval stage, as it overly depends on appearance-related knowledge. Therefore, we delve into a novel ReID problem with practical ramifications, namely short-term to long-term ReID (S2L-ReID), which continuously learns from ST to LT task. Existing continuous learning methods, only designed to mitigate domain differences across ST domains, face substantial challenges when dealing with the significant task gap in the S2L problem. These challenges manifest as severer catastrophic forgetting due to knowledge conflicts between ST and LT tasks and mutual constraints during optimization between over-coupling new knowledge adaptation and old knowledge non-forgetting tasks. To address these challenges, we propose a unified Meta-Knowledge Accumulation framework (MKA), which enables continuous learning of universal meta-knowledge for all tasks. It comprises a plug-and-play Meta-knowledge Purify Module (MPM) to alleviate knowledge conflicts by filtering out task-specific knowledge. Additionally, a transferable Peer Learning Strategy (PLS) is included to weaken mutual constraints by alternating between old and new models and enabling mutual learning. Finally, considering that real continuous application processes need to process the retrieval from ST to LT data, we propose a new Joint-Tests evaluation for evaluating the performance of the model on the more realistic hybrid data. Notably, MKA requires only a small computational resource usage with good portability. Empirical evaluations on all test settings exhibit that our MKA attains the best performance and is highly portable.