Selective Search Study Notes

selective Search
National Engineering Laboratory for Video Technology, Peking University
Diving csdn For many years, for the first time to write something, in the Peking University to do graduation design, do is about object parts detection, that is, from coarse to fine detection, the previous period began to learn RCNN related things, read paper, These days again selective search paper took out read, write some notes, talk about their understanding. The
Selective Search for Object recognition is a paper on recommendation area generation (proposal) published on IJCV, and the author is J.r.r.uijings, In this paper, a method called selective searching (selective search) is proposed for object recognition and detection. First involved in the detection of related topics, after looking at a few paper, found before many algorithms are based on Brute force search (exhaustive searching), the whole picture to scan, or the use of dynamic window method, this method of time-consuming serious, operation trouble. J.R.R proposed a selective search method, in the early identification of the whole picture in the generation of 1~3k proposal method, and then to each proposal processing. (i) Introduction

The

picture contains a variety of different levels of semantics and is informative, including textures (Texture), shapes, colors (color), and so on. The detection problem requires not only the recognition of the object in the image, but also the localization. According to author, when the object is identified, it is necessary to take full account of the diversity of the object (diversity), in addition, the objects hierarchical relationship (hierarchical) in the image is very important. The illustrations given in the paper are illustrated as examples.

Figure (a) shows a table with bowls on the table, a spoon in the bowl, and if the table is identified, the objects on the table need to be identified, and the layers between the different objects in the diagram are obvious. Figure (b) in two cats, can use color to separate them, but this time using the texture is not very effective; (c) On the contrary, the color of the environment of the chameleon and background is very similar, but if the texture is used, it is easy to separate the two, and in figure (d), the body and Tyre, There are a lot of differences in color and texture.
Overall, it is possible to understand the author's ideas:
1) The hierarchical information between images is important, and the search (exhaustive selective) adapts to different scales of objects by changing the window size, select search (selective Search) is also not able to avoid this problem. The algorithm uses image segmentation and uses a hierarchical algorithm (hierarchical algorithm) to solve this problem effectively.
2) A single policy cannot be applied to all images, the author proposes to use different strategies to merge the segmented images, for example: color, texture, size, etc. (ii) region merging algorithm

The author uses the hierarchical grouping algorithm to generate the initialized area, the specific reference paper "1", compared to the pixels, the region contains more information, the characteristics are more representational. The region merging algorithm is as follows:

1) Use FELZENSZWALB (2004) (Reference paper "1") to generate the initial segmentation area;
2) Initialize the similarity set S =∅;
3) Calculating the similarity between adjacent regions, and counting the results into s;
4) Merge the two region RI and RJ with the largest similarity into RT to replace RI and RJ, and re-compute the similarity again;
5) Get bounding box (iii) diversification strategy for each region (diversification strategies)

The author proposes three strategies in the selective search method:
1) by using a variety of color spaces with different invariant properties;
2) by using a different similarity measure;
3) by using a different initialization area;
In this article, for practical problems, the author uses a single color space, but in the similarity metric, different strategies are used to combine the methods:
1) Color similarity: S (color)
2) Texture similarity: S (texture)
3) Size similarity: S (size);
4) anastomosis similarity: S (FIT)
The specific calculation method refers to paper, and finally combines four metrics into one strategy: s = a*s (color) + b* s (texture) + c*s (size) + d*s (FIT);
The specific task of the project is to carry out a coarse-to-fine test of the car, accurate to the various components: rearview mirror, headlight, front windshield. Taking into account the selective search for the detection performance of the Widget object, a number of experiments were done to compare the different similarity measurement methods and the performance after the policy merging.

The above about the selective search generated proposal some of the principles and different strategies, the use of specific strategies also need to be based on practical issues fine tuning, from the table above, for the windshield and headlights proposal performance is very good, However, the performance of the rearview mirror is somewhat unsatisfactory, still need to continue to improve according to practical problems.
The entry is not deep, the above understanding inevitably have incorrect place, hope everybody correct, together progress, follow up will continue to update about Rcnn,fast Rcnn,faster RCNN and other related blog posts, hope to communicate with everyone.

Paper "1" Felzenszwalb p F, Huttenlocher D p. Efficient graph-based Image segmentation[j]. International Journal of Computer Vision, 2004, 59 (2): 167-181.