Transcript Document
Real-Time Compressive Tracking Gao Yuefang 2013.04.16 some slides are from Kaihua Zhang Outline • • • • • Paper & Author information Online Visual tracking Introduction Compressive Tracking Algorithm Experimental results Questions Paper & Author information • Paper: Real-Time Compressive Tracking, ECCV2012 • Authors: Kaihua Zhang PhD, Depart. Of Computing, the Hong Kong Polytechnic University http://www.comp.polyu.edu.hk/~cskhzhang/ Lei Zhang Associate Professor, Depart. Of Computing, the Hong Kong Polytechnic University http://www4.comp.polyu.edu.hk/~cslzhang/ Ming-Hsuan Yang Assistant professor, Electrical Engineering and Computer Science, University of California, Merced http://faculty.ucmerced.edu/mhyang/ Outline • • • • • Paper & Author information Online Visual tracking Introduction Compressive Tracking Algorithm Experimental results Questions Online Visual tracking Introduction • Invariant feature detectors (e.g. SIFT,SURF) • Online learning (e.g. Boosting, MIL) • Object detection (e.g. HOG for people detection) Online Visual tracking / Tracking by Detection Online Visual tracking Introduction (cont’) • One-shot learning for the first frame • learn and update appearance model during tracking Online Visual tracking Introduction (cont’) Online Tracking methods: • Generative methods learn a model to represent the object and then use it to search for the image with minimal reconstruction error. • Discriminative methods pose tracking problem as a binary classification task in order to find the decision boundary for separating the object from the background. • Collaborative methods Outline • • • • • Paper & Author information Online Visual tracking Introduction Compressive Tracking Algorithm Experimental results Questions Compressive Tracking Algorithm Existing Online Tracking methods often have the following problems: Online appearance models are data-dependent. As a result of self-taught learning, the mis-aligned samples are likely to be added and degrade the appearance model. How to solving the above problems? • appearance model based on data-independent, which employs nonadaptive random projections and preserves the structure of the image feature space. • a very sparse measurement matrix is adopted to efficiently extract the features for the appearance model Compressive Tracking Algorithm (con’t) Paper content: Appearance model based data-independent • Multi-scale image feature representation; • Random projection of above image features; Classifier • Naïve Bayes Classifier Appearance model and classifier updating Compressive Tracking Algorithm (con’t) Compressive Tracking Algorithm (con’t) Main Steps: Multi-scale image feature representation • The dimensionality of X is very high. Compressive Tracking Algorithm (con’t) How to choose the random matrix R? R should satisfy the following two properties: • Restricted isometry property(RIP)in compressive sensing theory: ensure the low dimensional features preserve the intrinsic structure of the high-dimensional features; • Very sparse: For computational efficiency; Compressive Tracking Algorithm (con’t) Appearance model using the following feature vector f Compressive Tracking Algorithm (con’t) Classifier and updating • Use Naïve Bayesian classifier • Online update scheme for parameters in the classifier Outline • • • • • Paper & Author information Online Visual tracking Introduction Compressive Tracking Algorithm Experimental results Questions Experimental results 实验视频段来自如下几个数据集: • MIL data set • VTD data set • L1 data set Experimental results(con’t) • David:能正确跟踪,仅有少许偏差,但不影响跟踪结果 #1 #276 #238 #303 #258 #319 Experimental results(con’t) • Bolt:无背景干扰正确跟踪,反之,完全跟丢 #1 #126 #153 #85 #136 #108 #149 #161 Experimental results(con’t) • Biker:能正确跟踪,但跟踪窗口大小固定,不能处理放大情况 #1 #43 #90 Experimental results(con’t) • kitesurf:未发生显著形变前,能跟踪,之后,不能跟踪到目标 #1 #30 #23 #31 #25 #34 Experimental results(con’t) • animal:基本跟踪不到目标 #1 #4 #12 Experimental results(con’t) • football:目标中度遮挡或完全遮挡时,跟丢目标 #1 #44 #62 #73 #54 #82 Outline • • • • • Paper & Author information Online Visual tracking Introduction Compressive Tracking Algorithm Experimental results Questions Questions • 图像多尺度表示: 其物理意义是什么?有无必要这样表示? • 投影随机性问题 如何保证跟踪的稳定性?【演示Biker和其他视频段】 • 压缩感知问题 这里没有涉及到压缩感知理论;仅是基于随机矩阵投影 Questions(con’t) • David:均匀分割目标区域方法与多尺度方法比较,跟踪效果没多大差 别,其中黄色为原始方法;其他很多视频情况类似,【运行该跟踪视频】 #1 #115 #4 #120 #94 #274 Questions(con’t) • Dollar、football:MSER与多尺度方法比较,跟踪稳定性获得稍许 改善,【演示该效果】。 实验结果总结 • 算法运算速度快,能实时跟踪 • 每个视频都有可能出现最好结果,但不是每次都 重现; • 大量视频存在定位不准的情况; • 尺度缩放问题未能解决; • 当前算法跟踪错误后无法纠正; • 剧烈运动视频大概率跟丢; • 遮挡问题未能解决; • 光照显著变化时大概率跟丢;