Robust Object Tracking with Online Multiple Instance Learning
read more
Citations
High-Speed Tracking with Kernelized Correlation Filters
Online Object Tracking: A Benchmark
Object Tracking Benchmark
Exploiting the circulant structure of tracking-by-detection with kernels
References
Rapid object detection using a boosted cascade of simple features
Greedy function approximation: A gradient boosting machine.
The Pascal Visual Object Classes (VOC) Challenge
A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting
Additive Logistic Regression : A Statistical View of Boosting
Frequently Asked Questions (13)
Q2. What are the future works in "Robust object tracking with online multiple instance learning" ?
One interesting avenue for future work would be to combine these ideas with the ones presented in this paper.
Q3. What is the main component of a typical tracking system?
A typical tracking system consists of three components: 1) an appearance model, which can evaluate the likelihood that the object of interest is at some particular location, 2) a motion model, which relates the locations of the object over time, and 3) a search strategy for finding the most likely location in the current frame.
Q4. What is the way to estimate the classifier weights?
Since the formulas for the example weights and classifier weights in AdaBoost depend only on the error of the weak classifiers, Oza proposes keeping a running average of the error of each hk, which allows the algorithm to estimate both the example weight and the classifier weights in an online manner.
Q5. What is the problem with adaptive appearance trackers?
In particular, if an object is completely occluded for a long period of time or if the object leaves the scene completely, any tracker with an adaptive appearance model will inevitably start learning from incorrect examples and lose track of the object.
Q6. What is the way to train the appearance classifier?
The authors argued that using Multiple Instance Learning to train the appearance classifier results in more robust tracking and presented an online boosting algorithm for MIL.
Q7. How does the algorithm choose the weak classifiers?
the algorithm sequentially chooses K weak classifiersfrom this pool by keeping running averages of errors foreach, as in [34], and updates the weights of h accordingly.
Q8. How many boosting algorithms have been proposed to learn the MILBoost model?
There have been many boosting algorithms proposed to learn this model in batch mode [39], [40]; typically this is done in a greedy manner where the weak classifiers are trained sequentially.
Q9. How do the authors model the instanceprobability of the weak classifiers?
The authors model the instanceprobability aspðyjxÞ ¼ HðxÞ ; ð8Þwhere ðxÞ ¼ 11þe x is the sigmoid function; the bag probabilities pðyjXÞ are modeled using the NOR model in (5).
Q10. Why does the OAB tracker perform better than the other two?
The authors acknowledge the fact that their implementation of the OAB tracker achieves worse performance than is reported in [25]; this could be because the authors are using simpler features or because their parameterswere not tuned per video sequence.
Q11. What is the main challenge of the adaptive appearance model?
A major challenge that is often not discussed in the literature is how to choose positive and negative examples when updating the adaptive appearance model.
Q12. How many positive images are generated per frame?
See text for details.r ¼ 1, generating only one positive example per frame (we call this OAB(1)); in the second variation the authors set r ¼ 4 as the authors do in MILTrack (although in this case, each of the 45 imagepatches is labeled positive); the authors call this OAB(45).
Q13. What is the way to crop out the image patches?
When updating the appearance model, the authors have the option of cropping training image patches only from the current scale or from the neighboring scales as well; in their current implementation the authors do the former.