"Summary" (CAS) Shambin teacher-"Overview of local image feature description"

"Summary" (CAS) Shambin teacher-"Overview of local image feature description"

This time we are honored to invite the Shambin of the Institute of Automation, CAS, to write our latest review of image characterization descriptors. Shambin has published several high-quality papers, including Tpami, PR, ICCV, CVPR, in the image feature description. His personal homepage is: http://www.sigvc.org/bfan/

We will continue to invite many teachers at home and abroad to do the latest Visual computing professional review report, such as feature extraction and description, sparse expression, human body tracking, three-dimensional clothing fabric animation, lightweight web3d, etc., and published in the Academic forum. Teachers will try to make the summary easy to understand, in simple, so whether it is a beginner or peer experts, will have some harvest. Because the academic articles published in this forum do not belong to the formal academic publication, the teachers will be able to organize the review into formal academic papers and publish them in the periodical conference according to the feedback from the forum.

In addition, if you feel that a scholar at home and abroad in your area of interest (in connection with visual computing: "Computer Vision", "graphics", "Pattern Recognition and machine learning", "robot vision navigation and positioning") do a good job, and hope to obtain his latest review of his/her contact information can be sent to us, We will consider inviting him/her to write the latest summary report so everyone can share it at the first time.
--------------------------------------------------------------------------------------------------------------- ---------------------------------

Overview of local image feature descriptionShambin
State Key Laboratory of Pattern recognition, Institute of Automation, CAS (Casia nlpr)

Local image feature description is a basic research problem in computer vision, which plays an important role in the search for the corresponding point in the image and the description of object feature. It is the foundation of many methods and therefore is also a hotspot in the current visual research, the annual top conference in the visual fieldICCV/CVPR/ECCVhave high-quality features to describe the publication of papers. At the same time, it also has a wide range of applications, for example, in the use of multiple two-dimensional images for three-dimensional reconstruction, recovery scene three-dimensional structure of the application, its basic starting point is to have a reliable image corresponding point set, and automatically establish the image between the point and point of the reliable correspondence between the two usually rely on a good local image Another example, in object recognition, one of the most popular and practical methods is based on local characteristics, because of the local characteristics, so that object recognition can deal with occlusion, complex background and other more complex situations.

the core problem of local image feature description is invariance (robustness) and distinguishable. Due to the use of local image feature descriptors, it is usually to deal with various image transformations in a robust manner. Therefore, in building/When designing feature descriptors, the problem of invariance is the first problem to be considered. In the wide baseline matching, it is necessary to consider the invariance of the variation of the angle of view, the invariance of the change of scale, the invariance of the change of rotation, etc. in shape recognition and object retrieval, it is necessary to consider the invariance of the shape of the feature descriptor.

However, the distinguishing character of the characteristic descriptor is often contradictory to its invariance, that is, a feature descriptor with many invariance, it is slightly weaker to distinguish the content of local image, and the robustness is often low if it is very easy to distinguish the characteristic descriptors of different local image content. For example, suppose we need to describe a fixed-size local image content around a point. If we directly expand the image content into a column vector to describe it, then as long as the local image content has changed a little, it will make its characterization of a large change, so that the feature description is easy to distinguish between different local image content, But for the same local image content rotation changes, and so on, it will also have a large difference, that is, the weak invariance.

on the other hand, if we use the Statistical Office image grayscale histogram for feature description, this descriptive way has a strong invariance, the local image content rotation changes and other situations are relatively robust, but the ability to distinguish weak, for example, can not distinguish two gray-scale histogram of the same but different content of the local image block.

to sum up, a good feature descriptor should not only have a strong invariance, but also should have a strong distinction.

in many of the local image feature descriptors,SIFT(Scale invariant Feature Transform) is one of the most widely used1999first proposed, to2004years to be perfected. SIFTis also a milestone work in the field of local image feature description sub-study. BecauseSIFTimage changes such as scale, rotation, and a certain angle of view and illumination change are invariant, andSIFTIt is highly distinguishable, and has been applied in object recognition, wide baseline image matching, three-dimensional reconstruction and image retrieval very quickly since it was proposed, and local image feature descriptors have been paid more attention in the field of computer vision, and a large number of local image feature descriptors have emerged.

SURF(speeded up robust Features) is theSIFTthe improved version, which leveragesHaarApproximate by waveletSIFTthe gradient operation in the method, and using the integral graph technique to calculate quickly,SURFthe speed isSIFTof the3-7times, most of the cases it andSIFTperformance, so it has been used in many applications, especially when the running time requirements are high.

DAISYIt is a fast computing local image feature descriptor for dense feature extraction, and its essence thought andSIFTis the same: the histogram of the statistical gradient direction of the chunking, the difference is that,DAISYin this paper, we improve the chunking strategy and use Gaussian convolution to collect the histogram of gradient direction, so the feature descriptors can be extracted quickly and densely by using the fast computation of Gaussian convolution. More coincidentally,DAISYThis feature aggregation strategy is by some researchers (Matthen Brown,Gang Hua,Simon Winderby means of machine learning, it is proved that it is optimal for several other feature aggregation strategies (block and polar coordinates under Cartier coordinates).

Asift(affine SIFTby simulating the images obtained from all imaging angles, we can deal with the change of view angle, especially the image matching under the change of large view angle.

Mrogh(Multi-support Region order-based Gradient histogramis the feature aggregation strategy to seek innovation, before the local image feature descriptor, its feature aggregation strategy is based on the location of the point in the neighborhood, andMroghfeature aggregation based on point-gray order.

BRIEF(Binary Robust independent Element FeatureThe local image feature descriptor is established by using the gray-scale relation of random points in the neighborhood of local image, and the two-valued feature descriptor not only has fast matching speed, but also has low memory requirement, so it has a good application foreground in mobile phone application. In fact, the idea of feature description using the gray-scale relation of point pairs in the neighborhoodSMD(ECCV ') has already been in it.

exceptBRIEF, many binary feature descriptors have been proposed in the last two years, for exampleORB,Brisk,FREAK. These feature descriptors are based on manual design, and some studies try to use machine learning methods to get the desired feature descriptors by data driven. Such feature descriptors includePca-sift,Linear discriminative Embedding,Lda-hashand so on. There are, of course, many other feature descriptors besides those described in the description, which are no longer described in one by one.

The most famous scholars in the international study of local image feature descriptors are:

United KingdomSurreyof the UniversityMikolajzyk, he wasInriawhen doing post-blogging, in the context of broad baseline application, theSIFT,Shape Context,Pca-sift, invariant moments and other local image descriptors are evaluated, and related papers are published in2005yearsPami, he proposed that the evaluation method is still a widely used performance evaluation method in the field of local image descriptor research.

Inriaof theC. Schmid, she began to study local image description methods in the 90 's, and was one of the elders in the field, but her team was shifting its focus over the years to large-scale image retrieval and behavioral recognition.

BelgiumLeuvenof the UniversityTinne Tuytelaars, she is famous forSURFthe author of the description,SURFrelated papers in .year obtainedCviuciting the most paper awards, she wrote three articles about local image characterization, respectively, "Local invariant Feature detectors:a Survey","Local Image Features"and"Wide Baseline Matching".

United KingdomOxfordof the UniversityAndrea Valida, he isvlfeatInitiator and principal author. vlfeatis an open source program that includesSIFT,Mserhas been widely used by many researchers. vlfeatother commonly used feature descriptors are being implemented gradually.

SwitzerlandEPFLof theVincent Lepetitand thePascal Fua, their team is focused on developing fast and efficient local image feature descriptors for template matching, three-dimensional reconstruction, and virtual reality applications. Their work includes the use of dense stereo matchingDAISYfeature descriptors, based onRandom Treestemplate matching method, based onRandom Fernstemplate matching method. In addition,Lda-hash,BRIEF,D-brief(ECCV) is also their masterpiece.

Wu Fuchao, a researcher at the Institute of Automation, CAS, has also done more in-depth research on this aspect, and has put forward many good local image feature extraction and description methods. These are the names that we often see when we read a paper.

the development trend of local image feature descriptors in recent years is: Fast and low storage. These two trends enable local image feature descriptors to play a role in fast real-time, large-scale applications, and to enable many applications to be developed on mobile phones, and to actually apply computer vision technology to the world around us. In order to satisfy the two requirements of fast and low storage, the two-valued feature descriptor has been widely concerned by researchers, whichCVPRand theICCVmost of the articles about local image feature descriptors are in this category. I believe they will continue to receive attention in the coming years and look forward to some successful applications in the public life.

"Summary" (CAS) Shambin teacher-"Overview of local image feature description"