QuickTopic (SM) free message boards QuickTopic (SM) free message boards
Skip to Messages
  Sign In to access your topic list  |New Topic |My Topics|Profile
Upgrade to Pro   Customize, show pictures, add an intro, and more:   QuickTopic Pro...and check out QuickThreadSM
Topic: Detecting and Reading Text in Natural Scenes
Views: 1010, Unique: 547 
Subscribers: 2
What's
this?
Printer-Friendly Page
Subscribe to get & post, or stop messages by email Subscribe
About these ads
Who | When
Messagessort recent-bottom   
Post a new message
 
Denzel  8
01-12-2007 09:43 PM ET (US)
Hello moder!
Paiton  7
07-21-2006 11:32 PM ET (US)
your discussion is very good! I will wait for updates. Give a look at allegra d side effects webpage devoted to allegra d side effects. lexapro medicine webpage devoted to lexapro medicine.
Hamed Masnadi-Shirazi  6
10-19-2004 03:31 AM ET (US)
Edited by author 10-19-2004 03:32 AM
They seem to have slower speed of detection compared to Viola Jones (that was able to detect faces at 30fps) even though they use less features and the same integral image trick. Why is that??

The features they used still seem similar to Viola Jones features, what is the difference? I did not understand their initial feature set production phase?
Robin Hewitt  5
10-19-2004 12:57 AM ET (US)
Their feature analysis process is interesting. I gather that the decisions about what category of feature to offer the classifiers at each level was largely done by hand. I wonder whether some of that process can be generalized and automated. For example, maybe good combination features can be discovered by examining the shape of their joint histograms to see how bimodal they are.

I didn't understand their modulus of derivative feature, though. What is that?
Andrew Rabinovich  4
10-19-2004 12:40 AM ET (US)
This is a good application for AdaBoost. However, the paper does not discuss the sensitivity of the algorithm to the input data and parameters. It would be interesting to see some tolerance bounds and the dependence of accuracy on the original signal.
Gary Tedeschi  3
10-19-2004 12:00 AM ET (US)
During the training of the strong classifier of the AdaBoost algorithm, the choice of weigths, alpha_t, and weak classifiers, h_t, are obtained by updating the weights, D_t(i). Why is the optimum choice that which minimizes Z_t ( where Z_t is the normalization factor)?
Stephen Krotosky  2
10-18-2004 04:41 PM ET (US)
This is interesting. I hope your talk goes deeper into the workings of Adaboost. Also, the authors make the claim that adaboost is the best training algorithm for detecting objects in image scenes. Where has this been investigated. It seems like depending on what features you choose, or how carefully you choose them, different training methods may have unpredictable results in terms of their comparative success.
Rasit Topaloglu  1
10-18-2004 04:25 PM ET (US)
Does anyone know how this "connected component algorithm" works?
RSS link What's this?
QuickTopicSM message boards
Over 200,000 topics served
Learn more Frequently asked questions  Acknowledgements
What they're saying about QuickTopic
 Questions, comments, or suggestions? Contact Us
Read our use policy before beginning. We value your privacy; please read our privacy statement.
Copyright ©1999-2006 Internicity Inc. All rights reserved.