Today, suddenly I felt that I had a lot of work to do, and then became sad (or might be depressed)!!! I should do my numerical optimization assignment by Friday (and I did not started it then), prepare a poster for Friday’s 603 final poster session, have a discussion with Mohammad about feature selection in RL, write my technical report for the Robotics challenges course, do my course project for the numerical optimization course, some possible coding for the Robotics Challenges course, prepare slides for CDC-ECC, write rebuttal for my previous paper, proofread the Memetic/Evolutionary/Learning/… paper, go shopping, and read many papers and books that I want to read, and while enjoying my life as my previous professor told me when I was going to come there! Worst of all, this NIPS 2005 will start next week in the middle of all these jobs! However, I do not want to miss that conference. It is a good experience for me to visit my colleagues.
Today we will have our second checkpoint for the Segway project. It should do something today, but I am pessimistic about its success.
In this project, I am involved in designing some controlling mechanism (moving from there to there) and detecting the drivability of a specific place. I do not like the second part.
HBD2SoloGen
Happy birthday to me! (:
Self-Organized Integration of Adaptive Visual Cues for Face Tracking
Jochen Triesch, Christoph von der Malsburg, Self-Organized Integration of Adaptive Visual Cues for Face Tracking
I have started reading a few papers on multi-cues and multi-modal hypothesis in visual object tracking. I guessed the field was not that interesting, but it seems that it is! The observer design problem in these situations is much more complex and realistic that what I have studied in my control courses. There, nothing was being said about multi-modality. Anyway, I summarize a few papers on this subject and express my ideas sometimes. Note that this weblog is semi-academic: so, my synopses may be not so accurate or correct. I would be happy if the author of papers help me understand their papers better. (:
It proposed a method called Democratic Integration. It is a weighted-based voting-like method that adapts the weights in a self-organizing manner, i.e. it does not use any external signal for those changes. Instead, it uses the difference between the overall result and the result of each cue. If they are alike, that cue’s weight (or reliability – as used in the paper) will be increased.
$$R(x,t)=\sum_i r_i(t)A_i(x,t)$$
$$\hat{x}(t) = argmax{R(x,t)}$$
And for adaptation, it defines a quality $q_i(t)$ as follows
$$\tilde{q_i}(t) = R(A_i(\hat{x}(t)) – E[A_i(x,t)])$$
in that $R(.)$ is the ramp function. In this formula, $A(.)$ is the saliency map of each tracker. It is stated in the paper and I see it in a few other that in the current literature, this saliency map is considered to be probability of the object to be in a specific place. The paper mentioned that this choice of quality function is ad hoc and some other ideas like using Kullback-Leibler distance would give better results.
utput of each tracker and represents After doing normalization over $q_i$, the change in reliability would be
$$\tau\dot{r_i}(t) = q_i(t) – r_i(t)$$
A similar formulation is given for adaptation of prototypes. The paper used a few simple cues to do face tracking. The result is not that exciting, but its performance comparing with the case without adaptation is much superior. I am not aware of the performance of other face tracking methods.
I wonder what would happen if I use context-based switching between different cues, i.e. storing the temporally steady-state reliability weights and then test them as initial guess whenever error happens. What is the measure of error?! I am not sure, but what about the number of conflicts between different trackers?! Or time-averaged gradient of reliability vector over time.
Muslims vs Jews
The long war between Palestinians and Israelians frustrates me. It is not fair to kill many people just for political disagreement between groups. I am against any radical solutions like erasing the Israel from the earth or throwing all Palestinians away from their homeland. I guess Muslims, Jews, and Christian can live with each other without any important and noticable conflicts. I am not a politician so I cannot analyze the politcal state and judge about the current situation very well (however, it seems that I’m not alone in this regard), but I just want to state my current opinion about the overal situation: let’s live with peace!
Aha … You may like to read this short piece of news.
Multi-modal Observers
Recently, I’ve become interested in multi-modal observers, e.g. particle filters. I like the idea of tracking different hypotheses simultaneously. In this viewpoint, we consider our estimation of the system’s state as pdf with arbitrary shape. In classical observers (e.g. Kalman filters), there is only a single unimodal hypothesis about the state. But in these particle filter-like methods, we consider different state estimations with different probabilities.
Well! This kind of viewpoint is not what was emphasized in my control theory courses. In those courses, everything were about proving the stability of the system (observer in this case) in a deterministic setting. We made no explicit attention to the probabilistic interpretation of the state estimation except that we assume that the measurment and state transition model are contaminated with a normal noise (so unimodal), and we evaluate the covariance of the state over time.
Now, I wonder what would happen to the control problem: Knowing the pdf of probable current states, what is the best control action? Suppose that one hypothesis says that your state is (10,0) and the other says that it is in (-10,0). If your controller is linear in states, what would you select? -K1*10 or -K1*(-10)?!
I must read more about these things. I am completely new in this field!
I’m ashamed of our president
Iran’s president told something which made horrible reactions in the world. I don’t know what is going on in Ahmadinejad’s head, but I know that if I like my country I would not talk that way. I guess he intentionally wants USA invade Iran!
SoloGen’s Adversarial Perturbation Method for Global Optimization
I want to propose an adversarial perturbation method for global optimization. I don’t know much about these perturbation methods for global optimization and I haven’t seen any theorem about them yet, but I guess this can be categorized as an adversarial perturbation method! I have not run any simulation yet since this morning that I found out that this method is as ridiculous as any other perturbation method! Anyway, I may do more investigations in the future:
Suppose that you can write your objective function in the additive form: f^k(X,\theta^k) = \sum^N_i(\theta^k_i g_i(X) ). The main objective function has all \theta^k_i equal to 1.
We are in step k and we have done a local search. So X^k is the local minimum of f^k(X,\theta^k).
Now as the perturbation, we can change \theta^k_i. If we change the sign of those \theta^k_i, we would get an “Alternating Trajectory Method” (named after whom?!). The conventional adversarial method tries to find the gradient of f w.r.t. theta and goes it uphill. I propose to change sign of a few \theta^k_i randomly with a probability of \mu in which \mu has the property that \theta^k_i will be 1 at the limit.
This way, we (hopefully) solve the primary objective function in the limit, we can change the minimum to a saddle point instantly (so, goes away from the local min very fast which is considered as an advantage of an adversarial method), and make a link between a trajectory method and a perturbation one.
Reviewing Night
Well! Just now, I finished writing three reviews for my Teaching and Research Methods course. Well! Those would not be the kind of reviews I like to do if they were real paper (I like to read cited papers too), but they are not that bad. Well! It is too late in the night, not eating any thing, and I should go home and cook something. Wow … a sample of hard working?!