The idea of significant target detection

Transfer from http://www.360doc.com/content/14/0725/09/10724725_396891787.shtml

1. Significant target detection introduction

significant testing has become a research hotspot in recent years, from the three major computer vision conferences (ICCV, CVPR, ECCV) on the number of articles can be seen, probably each session has 10 articles appearance, a so small topic,10 to the number of articles has been many. If you take a look at these articles, you will find that significant target detection accounts for a large part, and the eye movement Point predicts very little, about one or two articles. Seeing this, some people may not yet understand the difference between significant target detection and eye movement point prediction. In fact, the detection of significant targets is similar to a two-value segmentation problem, but it only adds a significant constraint, while the eye movement point prediction is to extract some points of interest to the human eye, rather than extracting a significant entire target area, that is, some pixels even on the significant target, it may not be attractive eye. So why are significant target detection articles so much in recent years? The reason, I personally think the first is this topic simple, do not need too deep mathematical theory, and do not need a physiological research base (eye movement point prediction more emphasis on the physical aspects of the human brain), easily out of the article, anyone can engage in; then, the application of the significant target detection is more directly in the computer vision, including image segmentation, namely saliency as Prior, to guide segmentation, to achieve unsupervised segmentation. Then, is the image classification, you can use saliency to improve the feature discrimination, such as the use of saliency to guide sparse code and so on, and finally, Daniel Itti and Borji and so on continue to push, Wrote a lot of benchmark articles, of course, more in the eye movement point prediction, there are ECCV12 about salient object detection benchmark. Now to send a top, it is best to and ECCV12 benchmark on the top ranked algorithm in the mentioned several database on the comparison. However, a topic sustained fever for some time, many people found that the detection of significant targets is becoming more and more difficult to do, no idea, can think of, have been used. So, I'll talk about how to proceed with the research of significant target detection.

2. Research ideas on the detection of notable targets

First, to learn from other related issues to solve the idea of doing saliency

Let me introduce the two most relevant issues with salient object detection:

1, segmentation

–figure/ground Segmentation or Matting –image Segmentation or Clustering –semantic Segmentation or image parsing

 2. Object Detection and recognition

–class-independent Object detection and recognition –class-specific Object Detection and recognition

Next, I'll give you some examples of how to get inspired from these related areas.

The first example is the ECCV12 Geodesic saliency, Msra's Yichen Wei, which mainly uses boundary prior and geodesic distances for significant testing. and similar ideas should be used in segmentation, such as geodesic Graph Cut for Interactive Image segmentation, CVPR 10, is to use and artificially labeled some of the prior areas of the geodetic distance, To measure the pixel is a target or background possibility, when we put the artificial calibration of the prior area, with boundary prior substitution, you can achieve bottom-up saliency detection. and boundary prior in constrained parametric min-cuts for Automatic Object segmentation, CVPR 10, has also been used, is to use boundary prior to guide the division.

The second example is CVPR13 salient Object detection:a discriminative regional Feature integration approach. Similar ideas, also in constrained parametric min-cuts for Automatic Object segmentation, CVPR 10, are used for segmentation. The difference is that one deals with region, and one is binary segment. The advantage of dealing with region is that it can be combined with various saliency mechanisms such as center-surround.

The third example CVPR13 saliency aggregation. The article mainly uses the local learning idea, that is, in-class differences, resulting in a model for the entire image library can not be good for each image, so he trained on each image of a local model. Similar ideas, in object recognition there are many, such as extracting Foreground masks towards Object recognition, ICCV 11. In addition, CVPR13 looking beyond image saliency also uses a similar idea.

Second, some new techniques are used to detect the significance of

We can borrow machine learning some technology, for example, sparse code,pca,manifold learning,ranking,graph model and so on to do saliency. These idea is easy to engage in, is to apply some technology to saliency come up. For a few examples, CVPR13 manifold ranking saliency, PCA Saliency, ICCV13 contextual Hypergraph modelling for salient Object Detecti On and so on.

3. Future research Ideas

I think the later salient object detection can do some things along several lines of thinking.

First, you can do some work in terms of speed, after all, salient object detection himself almost no use, is to do some advanced applications pre-processing.

Second, you can do some sparse code saliency research, seemingly sparse code in many aspects to fire a burst, such as classification,tracking,super-resolution, denoise and so on, is not in saliency also to fire a burst, seemingly sparse code effect is not very good. ICCV13 Huchuan Lu has an article sparse code, the effect is good.

Third, more use some prior to do saliency. For example, inter-image or out-image can be used to prior saliency this ill-posed problem into well-posed. This is a good direction, for example, we can learn statistical prior or find new prior, related ideas in segmentation also have, such as CVPRrobust region Grouping via Internal Patch Statistics and ACM MM10 Image segmentation with Patch-pair density priors.

The idea of significant target detection