How did you compute median from LK tracker? I assume LK will return velocity (u,v) ..Did you find the orientation from this (u,v) magnitude and find the median from this orientation ? For the trajectory- the P learning, did you predefined smooth spline equation for trajectory that object will follows? For the N learning , you said sth about local maxima. Did you compute the confidence weight from the updated model of median flow tracker (median LK?)
@ekalic2 How do you measure "robustness and speed" exactly? What are the units of measurement? Do "computer vision" people ever agreed on a method on how to do that? And LK is nothing but half of KLT. I'm sorry but it doesn't make it any better.
@ekalic2 Yes, of course. Nothing personal. What I'm dissatisfied with is the whole field of "computer vision". I think it's not science and that it is just a big bag of tricks borrowed from other fields and re-branded. Stuff like that. And I don't have to introduce myself. Think about me as a random encounter in Final Fantasy, except I lurk into the youtube comments sections.
@lennyhome Predator does not use KLT tracker. The fact that you don't know the difference between KLT and LK justifies your somewhat naive comment. With respect to failures. Sure, Predator is not the ultimate solution to all tracking problems. But if you show me a state-of-the-art tracker that outperforms Predator in terms of robustness and speed, I will give you 100$ ;)
@runbyway How do you do that? The answer is: you don't have to. In the computer vision field you make a speech and just pretend you can do it better than somebody else.
In other words this "Predator" is just the good old KLT with a fancy name, fancy features added, a rambling about "feeding back the error" and it fails pretty often outside of specially made demos. Am I right?
@coolman9999uk "pretty close to how we see and track objects with the human eye" I certainly don't do it that way, speak for yourself. Even animals are much smarter than that.
@ekalic2 In no other field it would be acceptable for a system to fail because a bird decided to perch somewhere nor it would be acceptable to present a product with no exact specifications and clearly stated operational limits expressed in centimeters, lumens or what have you. Something that can be proven or disproven, you know. And don't worry. I know you don't have any valid answer to the points I'm making. Nobody in computer vision does nor cares.
Is Google going to do a tech talk explaining why they ruined definition searches?
How did you compute median from LK tracker? I assume LK will return velocity (u,v) ..Did you find the orientation from this (u,v) magnitude and find the median from this orientation ?
For the trajectory- the P learning, did you predefined smooth spline equation for trajectory that object will follows?
For the N learning , you said sth about local maxima. Did you compute the confidence weight from the updated model of median flow tracker (median LK?)
Thats pretty good! Thats is open source project?
a bit late now, but yes :b
github.com/zk00006/OpenTLD
@ekalic2 How do you measure "robustness and speed" exactly? What are the units of measurement? Do "computer vision" people ever agreed on a method on how to do that? And LK is nothing but half of KLT. I'm sorry but it doesn't make it any better.
@ekalic2 Yes, of course. Nothing personal. What I'm dissatisfied with is the whole field of "computer vision". I think it's not science and that it is just a big bag of tricks borrowed from other fields and re-branded. Stuff like that. And I don't have to introduce myself. Think about me as a random encounter in Final Fantasy, except I lurk into the youtube comments sections.
@lennyhome Predator does not use KLT tracker. The fact that you don't know the difference between KLT and LK justifies your somewhat naive comment. With respect to failures. Sure, Predator is not the ultimate solution to all tracking problems. But if you show me a state-of-the-art tracker that outperforms Predator in terms of robustness and speed, I will give you 100$ ;)
@runbyway How do you do that? The answer is: you don't have to. In the computer vision field you make a speech and just pretend you can do it better than somebody else.
is it true that Predator became TLD, which became OpenTLD, which became CMT? They are all basically the same tracker, right?
Nope... TLD and OpenTLD are similar... CMT is completly different
In other words this "Predator" is just the good old KLT with a fancy name, fancy features added, a rambling about "feeding back the error" and it fails pretty often outside of specially made demos. Am I right?
this is cool for doing 3d with maybe.
Don't bother that guy @ekalic2 .. if you have time hope you can answer my question below
Terminators are close :D
@InPursuitOfALife Predator is open source already.
i think 0MoTheG is jealous of you Zdenek hihi
@runbyway The answers to your questions are discussed in my ICIP'10 and CVPR'10 papers. Please see my website (google zdenek kalal).
isn't this adaptive filtering? naming control-theory is rather confusing.
@lennyhome I'm sorry.. i'm not asking you.. i'm asking @ekalic2
@coolman9999uk "pretty close to how we see and track objects with the human eye"
I certainly don't do it that way, speak for yourself. Even animals are much smarter than that.
@ekalic2 In no other field it would be acceptable for a system to fail because a bird decided to perch somewhere nor it would be acceptable to present a product with no exact specifications and clearly stated operational limits expressed in centimeters, lumens or what have you. Something that can be proven or disproven, you know. And don't worry. I know you don't have any valid answer to the points I'm making. Nobody in computer vision does nor cares.
Is Google going to do a tech talk explaining why they ruined definition searches?