OpenTLD, anyone used it before? How does it work?

3
1

Has anyone used OpenTLD? It works by tracking an unknown bounded object and learns of the object in real time. There was a lot of buzz about this program a few months ago, and I'm not sure if I get all the hype for it. here is the homepage for it with video demos: http://info.ee.surrey.ac.uk/Personal/Z.Kalal/tld.html

Is this anything new and revolutionary? It seems to have many advantages compared to other classification algorithms, especially because it take no time to train and follows moving objects with relative ease. But what I'm confused about is, is this considered a object recognition algorithm or an object tracking algorithm?

If you look at the demos, every object required the user to create a bounding box on the object to be tracked. But is there anyway for it to work like a normal object recognition system and pick up the object it once learned?

asked Oct 05 '11 at 15:48

mugetsu's gravatar image

mugetsu
233●21●24●31

edited Oct 05 '11 at 16:33

Where was the "buzz" about it? I discovered it a couple of weeks ago because it was referenced by some elance projects.

(Oct 05 '11 at 20:39) finnw

@finnw I saw some articles, including on slashdot.

(Oct 06 '11 at 00:00) mugetsu
One Answer:
oldest newest most voted
2

I've only looked at the papers, not the code but as I understand it:

  • There is a classifier that looks at each region of the image and outputs a simple yes/no answer.
  • The "tracker" is an image registration algorithm that follows the object as it moves small distances per frame. The subimage at the new location/scale that it it predicts is fed into the classifier as a positive example.
  • For each frame, every region of the image (at multiple scales) is tested by the classifier. The best match (if the confidence is high enough) is output as the current location.
  • All matches in the frame other than the best one are fed into the classifier as negative examples

If any of these points makes it "revolutionary", it is the last one, because it balances the false negatives with the false positives.

The price paid for this is that it cannot track two or more similar objects in the same frame (showing it two of the original object will weaken it.)

answered Oct 05 '11 at 20:38

finnw's gravatar image

finnw
56●2●4

wow thats certainly a smart way of going around things. But doesn't doing this make the algorithm unable to properly work as an object recognition system? seems only applicable for cases where I want to track a specific object that I see in the video , not for picking up objects that I have learned in the past and telling my code what it is. agree?

(Oct 06 '11 at 00:09) mugetsu

@mugetsu yes but you could probably work around that by (1) keeping the model in a database or (2) Use the trick in the Predator YouTube demo where Zdenek shows it a card with pictures of possible matches and it picks out the same object that it has been tracking online.

(Oct 06 '11 at 08:36) finnw

are you by any chance on the google group? I think they would be able to give me some more details about this. Though it definitely is good news that Predator could potentially be modified for this purpose

(Oct 06 '11 at 14:02) mugetsu

@mugetsu, no I am not in the group

(Oct 08 '11 at 07:07) finnw
Your answer
toggle preview

learn more | privacy | contact

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.