Notes on Discriminative Tracking Feature - Bioinformatics I | CSE 598F, Study notes of Computer Science

Download Notes on Discriminative Tracking Feature - Bioinformatics I | CSE 598F and more Study notes Computer Science in PDF only on Docsity! CSE598G Collins On-Line Selection of Discriminative Tracking Features Robert Collins and Yanxi Liu (and later, Marius Leordeanu) ICCV 2003 CSE598G Collins Classification-based Tracking training frame test frame foreground background Classifier train a classifier label pixels F B B B CSE598G Collins Selecting Good Features for Tracking Motivation: real-time, adaptive feature selection for better distinguishing target from background while tracking. Approach: use a computationally simple method for computing “goodness” of each candidate feature so we can rank order them. “Goodness”  discrimination between foreground/background Blob Tracking Feature Selection CSE598G Collins Feature Selection Prior Work Feature Selection: choose M features from N candidates (M << N) Traditional Feature Selection Strategies •Forward Selection •Backward Selection •Branch and Bound Viola and Jones, Cascaded Feature Selection for Classification Bottom Line: slow, off-line process CSE598G Collins Prior Work Stern and Efros, Adaptive Color Space Switching for Face Tracking in Multi-Colored Lighting Environments, AFGR, 2002. RG rg HS CbCr Choose between five 2D color spaces Flesh probability image Pi formed by histogram backprojection using color space I Sample face and background windows selected Color space i evaluated using face Pi 2 Σ background Pi Σ||face|| CSE598G Collins Computing Variance Ratio variance of feature values on p={class1,class2} ratio of total variance (class1+class2 samples) to sum of variances of single class samples CSE598G Collins V ar ia nc e R at io higher lower Intuition Behind Variance Ratio CSES598G Collins feature y Related to Fisher Discriminant Function 0.8 0.6 0.4 o bo File: noname.mat, # of points K = 26 a im x tx | x L 1% 1 Ho ds L |! | ae 82 T rT? i T [ St rot Inter-Class Distance ee! yd L I iO H , re st) a or md? bolo ® D? + D3 6 0 1 a 90 + where m; = — S° yand D? = Yo (y—mi)?. © Ni yeC! yec! C4 ! ! ! ! ! L ! L -0.8 -0.6 -0.4 -0.2 0 02 0.4 0.6 08 feature x CSE598G Collins Computing Tuned Features empirical probability distributions weight image for tracking log likelihood ratio implementation detail ( avoid log(0) or xx/0 ) CSE598G Collins Example: 1D Color Feature Spaces (a R + b G + c B) (|a|+|b|+|c|) + offset where a,b,c are {-2,-1,0,1,2} and offset is chosen to bring result back to 0,…,255. Color features: integer linear combinations of R,G,B Note: this includes some common simple feature combinations R+G+B (intensity) 2G-R-B (excess green) R-B (opponent colors) Barring algebraically redundant features, we have 49 candidates CSE598G Collins Geometric Intuition on Color Candidate Features The 49 color feature candidates roughly uniformly sample the space of 1D marginal distributions of RGB. Color feature integer coefficient vectors displayed as unit vectors CSE598G Collins More Sample Feature Rankings Likelihood from most discriminative feature Object/background designation Likelihood from least discriminative feature CSE598G Collins Overview of Tracking Algorithm Note: since log likelihood images contain negative values, must use modified mean-shift algorithm as described in Collins, CVPR’03 Log Likelihood Images CSE598G Collins Avoiding Model Drift Problem: Adaptive appearance models have a tendency to “drift” background pixels mistakenly incorporated into the object model pull the model off the correct location, leading to more misclassified background pixels, and so on. Our solution: force foreground object distribution to be a combination of current appearance and original appearance (anchor distribution) anchor distribution = object appearance histogram from first frame object distribution = (current distribution + anchor distribution) / 2 Note: this limits the ability to drift, but also limits the ability of the appearance model to adapt to large color changes CSE598G Collins Benefits : Adapting to Changing Illumination / Background Trace of selected features CSE598G Collins Benefits : Minimizing Distractions For multi-color objects, the algorithm can avoid distractors by automatically adjusting color emphasis CSE598G Collins Benefits : Minimizing Distractions Current location Feature scores avoiding distractors by adjusting color emphasis CSE598G Collins A Tracking Failure CSE598G Collins Problem with Variance Ratio Although variance ratio does well at picking features that make the object appear distinctive from the overall background, it is unable to recognize spatially coherent clusters of high likelihood scores that represent potential distractors. As a result, mean shift may jump to a nearby vehicle. our target distractor tracking failure CSE598G Collins Problem with Variance Ratio Variance ratio examines the overall distribution of likelihood scores within the object region and surrounding background. Problem: this approach favors likelihood images with the object having high contrast with the average background score, even though there may be an equally high contrast distractor in the surrounding background region distractor higher score lower score weak object response but NO distractors We would prefer this image for tracking CSE598G Collins Distractor-Resistent Selection Solution: to avoid distractors, we must do better spatial reasoning about peaks in the location likelihood image. Our approach: 1) Given a candidate feature likelihood surface, smooth it with a Gaussian kernel related to the scale of the mean-shift window (the result represents the actual surface that the mean-shift algorithm performs hill-climbing on). 2) Extract the central object peak 3) Find the next highest peak – this represents the most likely distractor 4) Measure feature “goodness” as a function of these two peak heights (ratio or difference, depending on whether we are using a plain or log likelihood image) Note, this is related to the notion of “the margin” in a traditional classifier sense (in this case foreground/background classifier). We are thus choosing the feature that maximizes this margin, therefore also minimizing the probability of making a classification error. [we’ll see another reason why later] CSE598G Collins Example likelihood image distractors (after removal of target peak and rescaling) target peak worst distractor likelihood distractors likelihood distractors More Examples CSE598G Collins Peak Difference Computation Weight Image Primary Peak Object and Maximum (log likelihood) Identified Distractor Identified Peak Diff Score Prior Shape , = Estimate Weight Image Secondary Peak Identified CSE598G Collins Probabilistic Explanation Consider likelihood of two “events” We would like to maximize the likelihood of event A and minimize the likelihood of event B. We therefore want to maximize CSE598G Collins Probabilistic Explanation now crank through the math P(co = obj.c) = bg|.X0.%1) P(co = bg, c1 = obj|X0..%1) | apply Bayes rule, with priors denoted by 1 P(Xp.X1|Co = obj.cy = bg) N(cg = obj. c; = bg) P(X0,X1|co = bg, c1 = obj) (co = bg, ci = obj) | class conditional independence; replace constant prior by C c Palco = obj) P(Xiler = be) ~ P(Xolco = bg) P(Xi]ei = bj) CSE598G Collins Probabilistic Explanation P(Xo|co = obj) P(X |e; = bg) P(Xo|co = bg )P(X le1 = obj) | independence over pixels in region P(x|bg) ~ My Pi bar ll aos | substitute empirical distributions (computed from histograms) \ _ Dp ~~ (x) (x) x, 4 (x) "iG Pp  Top 3 Features, Variance Ratio Top 3 Features, Peak Difference (A) (B) (C) (D) (E) CSE598G Collins Segmenting the Object Observation: the likelihood image typically gives, visually, a reasonable description of the shape of the object. Make this concrete by segmenting out a binary shape mask Note: if we threshold likelihood at 0, we are actually making a color segmentation decision based on p(x|object) > p(x|background) for each pixel color x This assumes that, within our window, object pixels and background pixels are equally likely to occur. If that isn’t true, we should modify our decision rule to choose p(x|object)p(object) > p(x|background)p(background) CSE598G Collins Segmentation from Likelihood Image log likelihood threshold at 0 intersect region of interest region of interest CSE598G Collins Shape Comparison We have implemented shape comparison using based on chamfer distance Distance transform Add up values distance values where mask=1 shape 1 shape 2 score CSE598G Collins Comparison with Old Version old version (no shape) new version (incorporating shape) CSE598G Collins Comparison continued old version (no shape) new version (incorporating shape)