Regionlets for Generic Object Detection, regionletsgeneric

Regionlets for Generic Object Detection, regionletsgeneric

Regionlets for Generic Object Detection

This article is the translation and self-understanding of this article, the article: http://download.csdn.net/detail/autocyz/8569687

Abstract:

For general object detection, the current problem is how to use a relatively simple calculation method to solve the recognition problem caused by the angle change of the object. To solve this problem, a flexible object description method must be required, and this method can be used to evaluate objects in different positions.

Based on this situation, the author uses the cascade boosting classifier to establish an object classification model. This model contains different feature types. These feature types are obtained by calculating local regional features. These features are called regionlet.

Regionlet is the basic feature area, which is defined proportionally according to the detection window of any resolution. Place the regionlet with relative positional relationships in a group to depict the texture distribution of objects.

To adapt to the deformation of objects, the author combines those regionlets features into a one-dimensional feature. Then, the author calculates the border of an object, obtains the split start point through these borders, and limits the number of start points to thousands.

Introduction:

Although the detection of rigid objects (the shape will not change or the change is small) has achieved great success, but the detection of general objects still has many problems to solve. At present, the main problem is still the recognition problem caused by object deformation. There are two reasons for this problem. One is the deformation of an object, for example, a cat's. Different actions will change its appearance; another reason is that the visual angle and distance change. For example, although a vehicle does not change its appearance, it looks different from the angle and distance.

The problems mentioned above also indicate a very important issue for the expression of object types. On the one hand, a template that can well describe the characteristics of a rigid object may hardly be suitable for deformation objects. On the other hand, a template with good deformation and controllability may cause inaccurate positioning or relative error rates when detecting rigid objects.

 

In this article, the author proposes a new general object Expression Strategy, which combines adaptive deformation solutions into Classifier learning and feature extraction. In this Chinese method, the cascade boosting classifier is used to classify object frames. In boosting, each weak classifier uses the Regional Feature response in the box as the input, and these areas are expressed in a group of subareas in sequence. These subareas are called regionlets. Of course, these regionlets are not randomly selected, but are selected from a large candidate pool using boosting.

On the one hand, the relative location of regionlets in the region and the location of the region in the object box are relatively stable. Therefore, this regionlet expression method can establish a more detailed spatial expression model. On the other hand, the feature responses of each set of regionlets are merged into one-dimensional features, which have better robustness for local deformation.

In addition, in order to improve the flexibility of the regionlet model, the author adopted regionlet of different sizes and aspect ratios, and adopted selective search policies to obtain the number of candidate boxes in thousands of orders of magnitude, far less than the number of Sliding Window methods.

The main contributions of this article are as follows: 1. The regionlet method is proposed. This method can flexibly extract features from any box. 2. for a class of objects, based on the regionlet expression, not only is a relative spatial distribution model established within the object, in addition, by combining the regionlets selected by boosting and the feature response that aggregates a group of regionlets together, these two methods enable them to adapt well to the changes of objects, especially deformation.

 

Regionlet definition:

For object detection, object classification is defined by a classifier. This classifier contains the Appearance Features and spatial distribution of objects.

The appearance features of an object are generally extracted from the rectangular area of the object. Inside an object, a small rectangular frame is used to extract features. This feature has a good locality, but the deformation processing is poor. Although it has good processing capability for deformation, it cannot be accurately located by Using Large Rectangular frames. However, when an object changes significantly, especially when it is deformed, the large rectangular frame may not extract the body features. This is because some part of the information in the rectangle may be useless or even contain a margin.

In view of the above situation, the author wondered if he could find some sub-regions-regionlets, use these sub-regions as the basic template for feature extraction, and then put these templates in a group, such a set of features can more flexibly describe different objects, and have good controllability for deformation.

The figure above is used as an example. The first column in the figure is the object to be detected-the person, and the black box in the second column represents the size of the source image, the blue box inside is the area of feature extraction-R. Here, the extracted human features are mainly extracted from the upper half of the human body. The orange rectangle in the blue box is the sub-area of the feature extraction-regionlet. regionlet is the place where the picker is located, because although a person is deformed, his/her hand is less deformed. Combining the three regionlets in the second column becomes r1, r2, r3, or regionlets In the last figure.

The figure below is a careful analysis: the first is the selection of regionlet. Here we have selected representative people's hands. The bodies of the three figures are deformed, the biggest cause of deformation is the change in the hand position. However, although the hand position has changed, the hand deformation is relatively small in the figure. The selection of this set of regionlet is clever. Each individual regionlet is very representative and can highlight the characteristics of people, and is combined in a group of regionlets, we can accurately extract the features of our hands from different locations and handle the deformation well. Of course, this is just a simple one, in the actual algorithm, a region of R does not have only one regionlet, nor does it determine the location of the regionlet by manually analyzing features. As for how to determine these regionlets, I will talk about it later.

 

 

Feature Extraction in Region:

The process of extracting features from region -- R involves two steps:

Step 1: extract the HOG and HSV features of each regionlet respectively.

Step 2: Combine the features extracted from regionlets.

 

The first step is relatively simple. I will not repeat it here. I will explain in detail the implementation process of the second step.

The process of combining the features extracted by regionlet is actually a feature screening process. In regionlets, the author selects the items that best represent the region features.

For example, the author first extracts the series of low-dimensional feature features of these regionlets, obtains the learned dimension item, and then selects the most distinctive item through a boosting learning machine. It is found that the first item is the most distinctive, because in the regionlet area containing hands, its first item features are significantly higher than the first item of the other two regionlet. Finally, the author selects the one with the strongest feature response in the first item of the three regionlet as the feature expression of the entire region R.

Detection Window for Standardization:

The author's regionlet method is implemented in the object candidate box. To obtain the candidate box, refer to K. e. a. van de Sande, J. r. r. uijlings, T. gevers, and. w. m. smeulders. segmentation as selective search for object recognition. in ICCV, 2011. I will not repeat it here.

The method for obtaining the candidate box of appeal is used. After obtaining the candidate box, the system uses the detection window to detect the candidate area. Before the detection, the author normalized the detection window. The processing method is as follows:

As shown in, for figure (a), the image is a small candidate window with the size of (h, w). The current size is (l, t, r, B) detection box to detect this candidate window. When a large candidate window is created, assuming that the size is (B) corresponding to (h ', W'), if the size is (l, t, r, b) Obviously, the relative position of the detection changes, which does not conform to the character that the relative position of the regionlet remains unchanged. Therefore, the author first normalize the detection window in figure (, obtain the normalized proportional scale (l/w, t/h, r/w, B/h). When the detection window changes to a large (h ', W, the detection window is changed to (lw'/w, th'/h, rw '/w, bh'/h ). The normalized window method can be used for direct detection on images of different sizes.

 

Create region and regionlets pools:

The author has established a complete regiion and regionlet pool, which contains region and regionlet of different sizes, locations, and aspect ratios. The method is as follows:

In the method, R' = (L', T', R', B ', k) and k represent the K element of region's low-dimensional feature vector.