top bar
QuickTopic free message boards logo
New TopicNew Topic My TopicsMy Topics News
Skip to Messages

TOPIC:

Detecting and Reading Text in Natural Scenes

  Spam messages 10-9 deleted by QuickTopic 07-14-2011 08:17 AM
8
Denzel
01-12-2007
09:43 PM ET (US)
 Hello moder!
7
 Spam deleted by QuickTopic 07-14-2011 08:17 AM
6
Hamed Masnadi-Shirazi
10-19-2004
03:31 AM ET (US)
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?
Edited 10-19-2004 03:32 AM
5
Robin Hewitt
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?
4
Andrew Rabinovich
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.
3
Gary Tedeschi
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)?
2
Stephen Krotosky
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.
1
Rasit Topaloglu
10-18-2004
04:25 PM ET (US)
 Does anyone know how this "connected component algorithm" works?
Upgrade to PRO

Upload pictures, personalize your board, and more!

Print | RSS Views: 1529 (Unique: 863 ) / Subscribers: 2 | What's this?