在電腦視覺領域中，關鍵點偵測是理解物體結構的重要研究方向。這促使 了類別無關姿態估計（Category-Agnostic Pose Estimation, CAPE）這項新穎 任務的出現，該任務使用單一模型來定位不同物體類別的關鍵點。為了擴展其 應用，近期研究已將文字描述納入 CAPE 任務中。然而，將文字描述擴展至 CAPE 任務時，在骨架資訊學習和圖形資訊傳播方面仍存在不足，使得模型難以 充分利用結構資訊，進而影響關鍵點定位的精確度。在本論文中，我們提出了 一個整合語義骨架精煉器（Semantic Skeleton Refiner）並優化 Graph Transformer Decoder 架構的模型，利用文字描述的特徵作為動態結構預測的 引導。我們在 MP-100 資料集上進行實驗，並將我們的模型與 CAPE 領域中現有 的最先進模型進行比較。實驗結果顯示，我們的方法在 CAPE 領域的表現優於目 前的領先模型。 ;In the field of computer vision, keypoint detection is a crucial area of research for understanding the structure of an object. This has led to the emergence of Category- Agnostic Pose Estimation (CAPE), a novel task that utilizes a single model to lo- calize keypoints across diverse object categories. To broaden its application, recent research has incorporated text descriptions into the CAPE task. However, when extending text descriptions to the CAPE task, shortcomings in skeleton informa- tion learning and graph information propagation persist, making it difficult for the model to utilize comprehensive structural information, which in turn affects the precision of keypoint localization. In this thesis, we propose a model that integrates a Semantic Skeleton Refiner and optimizes the Graph Transformer Decoder archi- tecture, using features from text descriptions as a guidance for dynamic structure prediction. We conducted experiments on the MP-100 dataset and compared our model with existing state-of-the-art models in the CAPE domain. The experimental results indicate that our method outperforms current leading models in the field of CAPE.
