利用專家混合與域不變注意機制之域泛化行人重識別系統

Abstract

近年來，隨著對公共安全需求的增加以及監控系統的廣泛應用，行人重識別技術在相關研究中備受關注。儘管基於監督學習的方法在公開資料集上取得了顯著成果，但現實環境與訓練資料之間存在的領域差異，而且為每次佈署建立有標註的資料集是十分耗費人力的，所以這些方法並不能夠輕鬆地移轉至現實生活中的應用。

無監督域適應和完全無監督學習雖然在一定程度上解決了資料標注的問題，但它們往往會過度擬合於目標領域的風格，這對於多變的現實環境並不利。

鑒於上述問題，我們提出了一種基於域泛化學習的全新方法。該方法包括兩個關鍵模組：一是具備語義感知能力的注意力遮罩生成模組，二是專家混合與域不變注意層。前者有助於模型有效學習不同人體部位的特徵，後者通過域不變注意去除域相關資訊，並使用專家混合使模型使用不同參數處理不同數據，從而避免過度擬合於源域的分布

我們的方法在實驗中表現優異，超越了許多現有方法。這一創新方法將為行人重識別系統的實際應用帶來新的可能性，同時也推動了該領域的進一步發展。

In recent years, with the increasing demand for public safety and the widespread application of surveillance systems, person re-identification technology has received significant attention in related research. Although supervised learning-based methods have achieved remarkable results on public datasets, the domain discrepancy between real-world environments and training data poses a challenge. Additionally, creating labeled datasets for each deployment is labor-intensive, making it difficult for these methods to be easily transferred to real-life applications.

While unsupervised domain adaptation and fully unsupervised learning partially address the issue of data annotation, they often tend to overfit to the target domain style, which is disadvantageous for the dynamic nature of real-world environments.

Given the aforementioned challenges, we propose a novel method based on domain generalization learning. This method comprises two key modules: a Semantic Aware Mask Generator and a Mixture of Experts with Domain-invariant Attention layer. The former helps the model effectively learn features of different body parts, while the latter removes domain-specific information through domain-invariant attention and employs a mixture of experts to process different data with different parameters, thereby avoiding overfitting to the source domain distribution.

Our method performs exceptionally well in experiments, surpassing many existing methods. This innovative approach opens up new possibilities for the practical application of person re-identification systems and advances the field further.