Robust Tracking via Weakly Supervised Ranking SVM

参考文献：Yancheng Bai and Ming Tang. Robust Tracking via Weakly Supervised Ranking SVM

Abstract

通常的算法：utilize the object information contained in the current and previous frames to construct the object appearance model and locate the object with the model in frame t+1

问题：if the visual appearance fluctuates in short time intervals, suboptimal locations will be generated in frame t+1 if the visual appearance change substantially from the model.The continuous changes would accumulate errors and finally result in a tracking failure.

解决方法:online Laplacian ranking support vector tracker(LRSVT) to robustly locate the object; incorporates the labeled information of the object in the initial and the latest frames to resist the occlusion and adapt to the fluctuation of the visual appearance, and the weakly labeled information from frame t+1 to adapt to substantial changes of the appearance.

1、Introduction

困难：to resist the visual appearance changing frame by frame due to 3Drotation, sudden illumination changing and partial occlusion.

经典算法包含:image representation, the appearance model and the dynamic model. (appearance model plays a crucial role).

本文重点:focus on the model-free tracking problem, i.e., no prior knowledge except for the object location is known at the beginning of tracking.

其他人的算法：

1)construct and update the appearance model only using the current frame

2)the mean shift

3)learned a low-dimensional subspace representation of the object frame by frame, efficiently adapting to appearance changes.

4)learned a classifier as the appearance model via multiple instance boosting. The weak classifiers were online updated by means of a forgetting factor.

5)constructed the appearance model by means of linear and sparse combination of target templates and trival ones. The template set was dynamically updated according to the similarity between the tracking result and the template set.

6)used the initial frame and the latest four frames to learn the appearance model by means of sparse principal component analysis of a set of feature templates.

7)By introducing the local spasrse appearance model ,extended the mean shift based tracking algorithm and modeled the appearance in terms of a static sparse dictionalry of the object in the initial frame and an online updated histogram of the current frame.

8)formulated the tracking problem as a semi-supervised one,  where only the object bounding box in the first frame was considered as labeled, and all subsequent tracking results were letf unlabeled.(not good)

9)an online semi-supervised MILBoost tracker to combine the adptivity of multiple instance tracking and robustness aginst drifting of semi-supervised learning based tracking.

10)incorporated the positive and negative samples in the next frame to model the object appearance till the current frame, and located the object with the model in the next frame.

在substantially change下，这些模型不足。

文中的算法：

1)the target should be ranked higher than others around it

2)the relative relation between patches is easily figured out

3)extend the learning to rank algorithm, ranking SVM, to learn the relative relation.

4)provide rough loctions of the target object, i.e., some weakly labeled samples, in the next frame.

5)Based on above considerations, we propose a weakly supervised ranking SVM algorithm based on the smoothness assumption and the manifold regularization. called online Laplacian ranking suport vector tracking(LRSVT)

6)the labeled higher-ranked samples are composed of pathes very closed to the ground truth of the initial frame and those very close to the object locations in several most recent frames, and the labeled lower-ranked samples are those around the labeled higher-ranked ones. The weakly labeled higher-ranked samples are composed of patches close to that labeled as object patch by the weak labeler in the new frame.

7)These three sample sets are used to train the novel weakly supervised ranking SVM to result in the ranking function F(x).

8)then patches are sampled and ranked with the highest score by F(x) is accepted as the object location.

9)advantages:

　　incorporates the labeled information in the initial and the latest frames

　　incorporates the weakly labeled information in the next frame

2、Related Work

1)the topic of learning to rank, which combines relevance problems with prediction problems, has recently attracted considerable attention in machine learning community, and a great many of ranking algorithms have been proposed. The main goul of learning to rank is to automatically construct a ranking model based on the partial order of training data.

...

2)in the computer vision domian, learning to rank is mianly used in image and video retrieval.

3)learning to rank has begun to apply to other areas of computer vision. ..RankBoost...a two stage cascaded ranking SVMs detector

4)as a powerful machine learning technique, semi-supervised learning has been applied to cope with the visual object tracking problem.

3、Laplacian Ranking SVM

4、Tracking with Laplacian Ranking SVM

5、Experiments

6、Conclusions