Multi-instance learning (MIL) and Multi-pose Learning (MPL) are proposed by Boris Babenko at UC San Diego of cv. Their ideas can be summarized in the following figure. Mil refers to a learning instance of an object. There may be many situations in which learning is not based on a precise labeled object, instead, you can use a "package" composed of multiple instances of an object to learn. MPL means that an object has multiple pose ), when learning, it is often difficult to use a classifier to achieve good results. Therefore, you can train multiple classifiers to learn different pose. It describes the policy of simultaneous learning and alignment for an object in multiple situations, that is, mil is "adjusting training samples so they lie in corresponsible ", while MPL is "separating the data into coherent groups and training separate classifiers for each ".
On the right is the learning situation of mil, and each row is the learning data of an object; on the left is the learning situation of MPL, and the learning data of different pose of each behavior is, the frame of each color is a class for training. After the application of MIL was proposed, it seems that MPL has only a few references, but there is no specific formula for the MPL applied in the Multi-Cue onboard Pedestrian detection. Mpl differs from the traditional boost method by replacing the traditional y with Y in the following combination. k indicates multiple classes, that is, one class is recognized as 1, the result is 1.
The iterative Training Steps are also roughly the same. Apart from the complexity, there is one more layer for each YK training, and the corresponding update formulas are also different:
After that, Cambridge proposed multi-class learning (MCL), which is also used to train multiple classes. In addition to not using max (yk), it uses the following:
The probability model of Noisy-or, the overall thought is the same as that of MPL. However, MCL's iterative update method is much clearer. Its weight wki (k indicates each class, I indicates the sample) is updated using the following method:
Note the-1 ~ 1, but 0 ~ 1, so when a class is determined to be not of this type, this sample will be removed in the next training, and the corresponding judgment is that the weight of the class is increased. In a specific K class, the weight of the incorrect sample increases accordingly, which is in line with the traditional boost training method. MCL is used in data classification:
The effect looks pretty good ~
References:
[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, Robert to Cipolla. mcboost: Multiple classifier boosting for perceptual co-clustering of images and visual features
(Reprinted please indicate the author and Source: http://blog.csdn.net/xiaowei_cqu is not allowed for commercial use)