Which adaboost to use and when?

asked 2015-04-26 13:31:10 -0500

Tonystark124 gravatar image

If anybody has good prior knowledge about the adaboost types offered by opencv traincascade, DAB, RAB, LB and GAB, can you please explain which adaboost to choose and when to? I have started reading about them, but if anybody could shorten it, that would be of great help, for me and everyone who references this question. Thanks!

edit retag flag offensive close merge delete



I have an idea: if no one posts the answer, maybe you can post it, so it is also good for verification if someone else has another opinion... What you say?

thdrksdfthmn gravatar imagethdrksdfthmn ( 2015-04-27 04:54:12 -0500 )edit

This paper of Rainer Lienhart, creator of the tilted haar features approach, compares some of the available options and discusses their performance. Might be good to read this!

StevenPuttemans gravatar imageStevenPuttemans ( 2015-04-28 07:41:53 -0500 )edit

@thdrksdfthmn, deal. am already on it

Tonystark124 gravatar imageTonystark124 ( 2015-04-28 08:36:05 -0500 )edit

Real Adaboost has a good result in fast rotation invariant multi-view face detection. RAB mode in opencv_traincascade is the same with the real adaboost proposed by Bo Wu?

Sheng Liu gravatar imageSheng Liu ( 2015-10-19 03:02:36 -0500 )edit

The cascades are not rotation invariant ... so I am not sure how AdaBoost would lead to that rotation invariance.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-10-20 02:50:57 -0500 )edit

If he used rotation (or different poses) in training data, the cascade will detect the object even if rotation is present. I have to add that too much variation in training data will lead to bad cascades

thdrksdfthmn gravatar imagethdrksdfthmn ( 2015-10-20 03:49:13 -0500 )edit

That is ONLY true if you take objects that have a rotation invariant property like an overall shape that is round or squared. A cascade is on its own about -10 to +10 degrees rotation invariant, but thats it. Adding multiple orientations from 0 - 360 degrees just leads to a model that is so general that it will never detect the full object...

StevenPuttemans gravatar imageStevenPuttemans ( 2015-10-20 04:03:52 -0500 )edit

Thanks a lot!

Sheng Liu gravatar imageSheng Liu ( 2015-11-10 04:42:57 -0500 )edit