Main idear
Treat the tracking problem as a classification task and use online learning techniques to update the object model
Main Innovative Points
1 Based on structured output prediction (Structured SVM), in which the task is directly predict the change in object location between frames, instead of relying on a heuristic intermediate step for producting
labelled binary samples with which to update the classifier, which is often a source of error during tracking
2 The online structured output SVM learning framework is also easy to incorporate image features and kernels, and SVM also has good generalization ability,robustness to label noise and flexibility in object representation
through the use of kernels
3 不是採用离目标近的为正样本，远的为负样本这样的准側，而是採用overlap。大于一定值的为正，小于为负
4 核函数
5 对支持向量进行限制，将权重影响小的去掉
The issues raised by other tracking-by-detection approach
1 the classification confidence function provides an accurate estimate of object position is not explicitly incorporated into the learning algorithm, since the classifier is trained only with
binary labels and has no information about transformations
2 Examples used for training the classifier are all equally weighted, meaning that a negative example which overlaps significantly with the tracker bounding box is treated the same as one which
overlaps very little. One implication of this is that slight inaccuracy during tracking can lead to poorly labelled examples, which are likely to reduce the accuracy of the classifier, in turn leading to further tracking inaccuracy
3 the labeller is usually chosen based on intuitions and heuristics, rather than having a tight coupling with the classifier. Mistakes made by the labeller manifest themselves as label noise,
and many current state-of-the-art approaches try to overcome this problem by using robust loss functions [13, 14], semi-supervised learning [11, 17], or multiple-instance learning [3, 23]. We argue that all of these techniques, though justified in increasing the
robustness of the classifier to label noise, are not addressing the real problem which stems from separating the labeller from the learner
How To Do
1 总览
算法分为两步：1 预估物体的位移 2 更新判别函数
Structure Learning是一种同意输出为结构的学习方法。理论上不论什么输出都能够作为一种结构。即能够解决随意问题
Structure SVM 是结构学习的一种，已是一种比較成熟的算法实现框架，论文參考參考附件部分1,2,3 ，代码网址
2 建立 Structure SVM 模型
文中的模型为：
当中的约束条件是从>=0进化来的
1 >=0时w的解不唯一。所以我们选择是间隔最大的w并限制w的长度，——> >=1
2 松弛>=1-
3 将1换为损失函数，越不同，间隔要求越严格（大）
3 解这个SVM模型
A
核心步骤基于SMO（序列最小最优化）Style的，SMO的基本思路是：选择两个变量（至少一个违反KKT）。固定其它变量，进行两个变量的二次规划问题求解，这样将问题不断的分解为子问题进行求解，进而达到求解原问题的目标。SMO參考4,5
B
文中还引入了Budget来对支持向量的个数进行约束。进而能够达到实时方法为：
Similar to [21], we choose to remove the support vector which results in the smallest change to the weight vector w, as measured by ||∆w||2
C
Search over Y on a polar grid rather than considering every pixel offset.
实验
1 採用的2个尺度的4X4的6种不同的haar-like（192 features）
2 Combine some different features by averaging multiple kernels
參考附件
1 Large Margin Method for Structured Learning
2 Support Vector Machine Learning for Interdependent and Structured Output Spaces 1的简短版
3 Structured Learning and Prediction in Computer Vision
4 统计学习方法-李航 7.4
5 Sequential minimal optimization: A fast algorithm for training support vector machines