The paper "Struck: Structured Output Tracking with Kernels" by Sam Hare, Stuart Golodetz, Amir Saffari, Vibhav Vineet, Ming-Ming Cheng, Stephen L. Hicks, and Philip H. S. Torr presents a novel framework for adaptive visual object tracking. The authors address the limitations of traditional tracking-by-detection methods, which often separate the learning and tracking processes, leading to issues such as inefficient sample generation and label noise. Their proposed framework, called Struck, integrates learning and tracking by using a structured output support vector machine (SVM) to directly predict the change in object configuration between frames. This approach avoids the need for an intermediate classification step and explicitly couples the classifier's goal with the tracker's goal. Struck uses a kernelized SVM, which can be computationally expensive due to the "curse of kernelization," where the number of support vectors can grow unboundedly. To address this, the authors introduce a budgeting mechanism that limits the number of support vectors, ensuring efficient evaluation and real-time tracking. They also implement Struck on a GPU to achieve high frame rates, making it suitable for real-time applications. The paper includes a detailed description of the Struck algorithm, including its primal and dual SVM formulations, and discusses various practical considerations such as parameter vectors, search spaces, kernel functions, and image features. Experiments demonstrate that Struck outperforms state-of-the-art trackers on benchmark datasets, and the authors show how to incorporate additional features and kernels to further improve tracking performance.The paper "Struck: Structured Output Tracking with Kernels" by Sam Hare, Stuart Golodetz, Amir Saffari, Vibhav Vineet, Ming-Ming Cheng, Stephen L. Hicks, and Philip H. S. Torr presents a novel framework for adaptive visual object tracking. The authors address the limitations of traditional tracking-by-detection methods, which often separate the learning and tracking processes, leading to issues such as inefficient sample generation and label noise. Their proposed framework, called Struck, integrates learning and tracking by using a structured output support vector machine (SVM) to directly predict the change in object configuration between frames. This approach avoids the need for an intermediate classification step and explicitly couples the classifier's goal with the tracker's goal. Struck uses a kernelized SVM, which can be computationally expensive due to the "curse of kernelization," where the number of support vectors can grow unboundedly. To address this, the authors introduce a budgeting mechanism that limits the number of support vectors, ensuring efficient evaluation and real-time tracking. They also implement Struck on a GPU to achieve high frame rates, making it suitable for real-time applications. The paper includes a detailed description of the Struck algorithm, including its primal and dual SVM formulations, and discusses various practical considerations such as parameter vectors, search spaces, kernel functions, and image features. Experiments demonstrate that Struck outperforms state-of-the-art trackers on benchmark datasets, and the authors show how to incorporate additional features and kernels to further improve tracking performance.