"Pattern recognition" mpl,mil and MCL

Multi-instance Learning (MIL) and Multi-pose Learning (MPL) are CV's Daniel Boris Babenko at UC San Diego, whose ideas can be summarized in the following picture. Mil refers to a learning instance of an object that may be in many cases, not learning with a precise labeling object, but rather using a "package" consisting of multiple instances of an object, and the MPL refers to an object that has multiple gestures (Pose), When learning to use a classifier is often difficult to achieve good results, so you can train multiple classifiers to learn different pose. It describes a strategy for learning and aligning multiple situations of an object, that is, the mil is "adjusting training samples so they lie in correspondence" and the MPL is "separating the data into coherent groups and training separate classifiers for each ".

The right side of the picture above is the study of the mil, the learning data of each action object; The left is the MPL learning situation, each behavior of an object's different pose learning data, and each color box is a class of training. Mil was put forward for a wide range of applications, and seemingly only a few of the MPL, an article "Multi-cue Onboard pedestrian Detection" application of the MPL but there is no specific formula. The difference between the MPL and the traditional boost method is to use the following combination of Y instead of the traditional Y, where k represents multiple classes, that is, a class is recognized as 1, then the result is 1.

Its iterative training steps are also roughly the same, in addition to the complexity of a layer of each YK training, in addition to the corresponding update formula is also different:

Cambridge then proposed Multi-Class Learning (MCL), also training multiple classes, except for the use of Max (YK) instead of using the following:

The probability model of noisy-or, the whole sense of thought and the MPL is the same. However, MCL's iterative Update method is much clearer to write, its weight wki (k means each class,i represents a sample) update using the following methods:

Pay attention to the common -1~1 in the P-class model, but 0~1, so when a class is judged not to be this class, the sample will be removed in the next training session, and the corresponding judgment is that the weight in the class is increased. In the specific class of K, the weight of the sample will be increased correspondingly, which fits the traditional method of boost training. MCL's article is used in the data classification above:

The effect looks very delicious ~

Reference documents:

[1] Boris Babenko, Piotr Doll´ar, Zhuowen Tu, Serge Belongie. Simultaneous learning and Alignment:multi-instance and multi-pose learning.

[2] Christian Wojek, Stefan Walk, Bernt Schiele. Multi-cue Onboard Pedestrian Detection

[3] Tae-kyun Kim, Roberto Cipolla. Mcboost:multiple Classifier boosting for perceptual co-clustering of Images and Visual Features

(reprint please indicate author and source: Http://blog.csdn.net/xiaowei_cqu do not use for commercial use without permission)