近年來，基於孿生網路（Siamese networks）之追蹤方案，大多採用互相關（cross-correlation）計算目標物模板與搜索畫面中各個區域的相似度，並透過分類（classification）網路與迴歸（regression）網路分別預測目標物之位置與邊界框（bounding boxes）之座標。然而，由互相關產生之分數圖（score map）僅能大致呈現目標物的所在位置，無法精確反映出目標物的主要語意特徵，而分類網路與迴歸網路之間缺乏交流機制，導致分類結果無法正確反映網路所預測的邊界框準確性。因此，本論文提出基於注意力機制（attention mechanism）以及殘差連接（residual connection）的特徵強化模組，並進而應用於基於孿生網路的物件追蹤器之單向與雙向（bi-directional）的特徵強化，其中，單向模組用於取代互相關運算，使追蹤器得以利用具有語意訊息之特徵進行更準確的邊界框預測，雙向模組則用於分類網路與迴歸網路產生之特徵映射（feature embedding）相互進行聚合與強化，使兩者能夠交流資訊，並於訓練階段能間接由彼此之損失函數（loss function）輔助學習。本論文於大型追蹤平台GOT-10k及LaSOT進行測試，實驗結果顯示所提出之追蹤器相較於最先進之方案，在短期與長期追蹤上能兼顧準確率與追蹤速度（67 FPS）。;In recent years, cross-correlation has been used in most Siamese-based trackers for similarity measuring between a target template and a search region, where a classification network and a regression network are adopted for target localization and bounding box prediction, respectively. However, the score map generated by cross-correlation can only approximate the target location, failing to represent semantic information of the target. The lack of communication mechanism between the classification and regression networks results in the misalignment between the classification results and the precision of the predicted bounding boxes. Thus, this paper proposes an attention mechanism based module with residual connection for unidirectional and bi-directional feature enhancement in Siamese-based trackers. The unidirectional module is used to replace cross-correlation, making the trackers able to predict more precise bounding boxes with semantic information. The bi-directional module aggregates and enhances the feature embedding generated by both classification and regression networks reciprocally, hence the two networks can exchange information and be optimized indirectly with the loss functions of each other during the training phase. Experimental results on benchmarks including GOT-10k and LaSOT show that the proposed scheme has balance between tracking accuracy and speed (67 FPS) compared to state-of-the-art trackers on both long-term and short-term tracking.
